Privacy-preserving AI-enabled video surveillance for social

distancing: responsible design and deployment for public

spaces

Author
Sugianto, Nehemia, Tjondronegoro, Dian, Stockdale, Rosemary, Yuwono, Elizabeth Irenne

Published
2021

Journal Title
Information Technology & People

Version
Accepted Manuscript (AM)

DOI
https://doi.org/10.1108/ITP-07-2020-0534

Copyright Statement
© 2021 Emerald. This is the author-manuscript version of this paper. Reproduced in accordance
with the copyright policy of the publisher. Please refer to the journal's website for access to the
definitive, published version.

Downloaded from
http://hdl.handle.net/10072/407130

Griffith Research Online

https://research-repository.griffith.edu.au
 Information Technology & People

Inf
or
Privacy-preserving AI-enabled video surveillance for social
distancing: Responsible design and deployment for public
spaces
m

Journal: Information Technology & People

at
Manuscript ID ITP-07-2020-0534.R2
io
Manuscript Type: Article

Surveillance < Phenomenon, Information management < IT/IS

nT
management < Practice, Ethics, Agile computing < Technology,
Keywords:
knowledge transfer < Phenomenon, Adoption < Information system
development < Practice
ec
hn
olo
gy
&
Pe
op
le
Page 5 of 44 Information Technology & People

1 Privacy-preserving AI-enabled video surveillance for social distancing:

2 Responsible design and deployment for public spaces
3
4
5
6 Abstract
7 Purpose – The paper proposes a privacy-preserving artificial intelligence-enabled video surveillance technology to
8 monitor social distancing in public spaces.
9
Inf
Design/methodology/approach – The paper proposes a new Responsible Artificial Intelligence Implementation
10
11 Framework to guide the proposed solution’s design and development. It defines responsible artificial intelligence criteria
12 that the solution needs to meet and provides checklists to enforce the criteria throughout the process. To preserve data
privacy, the proposed system incorporates a federated learning approach to allow computation performed on edge
or
13
14 devices to limit sensitive and identifiable data movement and eliminate the dependency of cloud computing at a central
15 server.
ma
16 Findings – The proposed system is evaluated through a case study of monitoring social distancing at an airport. The
17
results discuss how the system can fully address the case study’s requirements in terms of its reliability, its usefulness
18
19 when deployed to the airport’s cameras, and its compliance with responsible artificial intelligence.
20 Originality – The paper makes three contributions. First, it proposes a real-time social distancing breach detection
tio

21 system on edge that extends from a combination of cutting-edge people detection and tracking algorithms to achieve
22 robust performance. Second, it proposes a design approach to develop responsible artificial intelligence in video
23 surveillance contexts. Third, it presents results and discussion from a comprehensive evaluation in the context of a case
nT

24 study at an airport to demonstrate the proposed system’s robust performance and practical usefulness.
25
26
27
Keywords artificial intelligence, responsible AI, data privacy, social distancing, video surveillance technology
Paper type Research paper
ec

28
29
30 1. Introduction
hn

31 COVID-19 is a global pandemic that spreads quickly through close human contact with infected patients. The
32 virus had infected 10 million people in June 2020 and 38 million people by October 2020 (World Health Organization
33 [WHO], 2020). Despite the economic impact of lockdown, government-imposed social distancing in public spaces is
34 critical for slowing the infection rate (Department of Health, 2020). Monitoring compliance of social distancing in public
olo

35
spaces is usually done by the organization (government or business entities like shopping center management). Manual
36
37 monitoring through video surveillance to identify breaches is arguably the least invasive method, but tedious and should
38 be automated. Artificial Intelligence (AI) can automate analysis of crowd density and detect social distancing breaches.
gy

39 When crowd density becomes critical and there are too many occurrences of people failing to maintain safe distances,
40 the automatic system would immediately trigger a notification to authorized personnel to take action. By observing
41 long-term patterns of crowd density, movement and spatial distancing, local government and business owners can also
42 gain insights for redesigning their space to better manage social distancing (Boland et al., 2020). AI can enhance the
&P

43
usefulness of existing video surveillance systems already installed in public spaces, and can be integrated as part of a
44
45 smart city’s intelligent infrastructure to maintain safe distancing and prevent COVID-19 spread (Feldstein, 2019).
46 Video surveillance raises data privacy concerns when humans are the interest of observations. Extensive use of
47 such technology can threaten people’s privacy and cause harm to people. Most existing AI-enabled video surveillance
eo

48 systems rely on centralized processing, meaning data are transmitted to a central/cloud machine for video analysis
49 (Kavalionak et al., 2019). Such an approach involves data privacy and data security risks (Sanchez-Iborra and Skarmeta,
50 2020). Serious data threats such as data theft, eavesdropping or cyber-attacks, can potentially occur during data
ple

51
transmission. Likewise, when AI-enabled video surveillance systems incorporate AI continual learning to adapt the AI
52
53
models towards constant visual change in the environment, existing approaches require new data to be collected from
54 cameras and retained long term (Sugianto et al., 2019). This also involves data privacy concerns due to potential data
55 misuse. Existing laws in Asia-Pacific countries require data to be de-identified or deleted immediately after use
56 (Kennedys Law, 2018). While many countries have established privacy laws to protect their people (Clifton et al., 2020),
57 such emerging technology still lacks the much-needed people’s and community trust as it is often considered technology
58 that invades their privacy. When people’s trust is lacking, AI solutions are less likely to be adopted regardless of the
59 intended benefits (Rossi, 2018).
60
This paper proposes a privacy-preserving AI-enabled video surveillance technology to monitor social distancing
in public spaces. To preserve data privacy, the proposed system incorporates a federated learning approach to allow
computation performed on edge devices rather than transmitting sensitive and identifiable data to the central machine.
 Information Technology & People Page 6 of 44

The paper makes three contributions. First, it proposes a real-time social distancing breach detection system on edge
1
that extends from a combination of cutting-edge people detection and tracking algorithms to achieve robust performance.
2
3
Second, it proposes a design approach to develop responsible AI in video surveillance contexts. The approach is guided
4 by a new Responsible AI Implementation Framework (RAIFF) to enforce responsible AI criteria throughout designing,
5 developing, and deploying the system. Third, it presents results and discussion from a comprehensive evaluation in the
6 context of a case study at an airport to demonstrate the proposed system’s robust performance and practical usefulness.
7
8 Case Study
9 The proposed AI solution was designed based on a case study at an airport which has 6.5 million (pre-COVID-19
Inf
10 impacts) passengers annually with 17,000+ passengers daily. Hundreds of integrated surveillance cameras were installed
11
to cover 290,000 square metres with hundreds of shops and 40+ check-ins. To ensure all passengers have a good
12
experience during their visits, the airport employs some trained staff to observe people’s activities across all areas
or
13
14 through surveillance cameras to ensure that the airport’s activities run smoothly, and the airport provides good service
15 to passengers. When a problem is noticed, they will notify other staff to respond immediately. For example, when the
ma
16 number of people in a certain area is above a safe number, in-field staff will be notified and asked to respond immediately
17 to ensure people’s safety. Given the complexity of such environments, it is humanly impossible to continuously observe
18 the airport’s activity just by relying on human effort. Existing human resources are limited in constant manual
19
observation given the large number of cameras to observe. Human distraction and fatigue may result in missing some
20
tio

21 important events while observing cameras.

22 The intelligent AI-based video surveillance system can help airport personnel observe people’s activities. The
23 AI solution will provide prompt airport activities analysis to the airport’s management in real-time. Authorized staff
nT

24 will receive automatic notifications when there is something urgent or critical and needs immediate follow up by
25 personnel. The use of such intelligent video surveillance technology is proposed as we consider existing available
26 resources at the airport given that surveillance cameras and the associated infrastructure are already well established.
27
Data privacy is the main concern as the proposed system will constantly observe people’s activities. While the
ec

28
29 airport realizes the benefits of such an intelligent system, they have consistently raised data privacy concerns since the
30 beginning of the project. The proposed system is expected to be able to deliver the benefits while not compromising
hn

31 passenger privacy and obeying existing laws and human ethics. As initial actions, strict agreement has been made to
32 protect any sensitive data such as people’s identities. People’s faces must be blurred and videos can only be stored up
33 to 7 days. The main goal is to develop an intelligent video surveillance while ensuring data privacy protection.
34 Adopting AI solutions at the international airport has evolved over time alongside addressing the system’s AI
olo

35
responsibilities. Prior experience with the AI solution sets precedence for the airport’s technology manager to trust the
36
37
development of the new capabilities. Table I summarizes the evolution of an AI solution at the airport, starting from
38 people detection and tracking to social distancing breach detection. In 2017, we initially developed automatic people
gy

39 detection and tracking to investigate the feasibility of the proposed intelligent system. Since people are the observation’s
40 point of interest at the airport, detecting and tracking people is crucial to establish the foundation of the system prior to
41 further system development. In 2018, we extended the system to automatically count crowd levels to ensure the number
42 of people in certain areas was below safe numbers. During the COVID-19 pandemic in 2020, the Australian government
&P

43 enforced businesses to conduct strict safety responsibilities such as practicing social distancing during business
44
operation. In response, the airport must rearrange its space to ensure passengers have safe distances between them during
45
46 their visit. Space rearrangement is mostly applied to critical areas, such as waiting areas and queue lines, where people
47 will gather in large numbers. Thus, it motivates us to extend the system further to detect any social distancing breaches
eo

48 and generate a heat map for spatial analysis purpose. By having such information constantly, the airport can understand
49 how well their social distancing practice works, evaluate and plan to rearrange their space to better ensure people’s
50 safety.
ple

51
52 ---Insert Table I here----
53
54
2. Related Work
55
56 Existing solutions to help to monitor social distancing among people can be categorized into wireless, sensory
57 signals and vision-based tracking devices. Both wireless and sensor technologies can only work if everyone within the
58 observation range uses the same device or tag and cannot present optimum performance without people’s cooperation
59 in enabling the signal transmission. Wireless location-based technologies track individuals by picking up signals from
60 their electronic devices, such as mobile phones, to get their location. Apple and Google recently released mobile
applications intended for tracing COVID-19 infected people using Bluetooth signals (Romm et al., 2020). Aside from
Bluetooth, other signals such as Wi-Fi Received Signal Strength Indicator (RSII), cellular, Ultra-WideBand (UWB),
Page 7 of 44 Information Technology & People

Global Navigation Satellite System (GNSS), and Zigbee signals are potential solutions for social distancing verification
1
(Al-Suwaidi and Zemerly, 2009, Laoudias et al., 2018, Ledergerber and D’Andrea, 2020, Mazuelas et al., 2009, Niu et
2
3
al., 2015). Sensor technologies pick up ultrasound, inertial, visible light or thermal frequency signals from specific tags
4 attached to the users (Jovicic and Engineer, 2016, Kok et al., 2015, Szajewska, 2017, Tiemann and Wietfeld, 2017). The
5 signal is then calculated using mathematical models to alert the users when social distancing is breached (Nadikattu et
6 al., 2020). In contrast, vision-based technology uses computer vision to automatically track people’s movements
7 without requiring any devices attached to each individual.
8
9 2.1. AI Vision Algorithms for Monitoring Social Distancing
Inf
10 The current challenge in existing AI-enabled vision-based people detection and tracking is to produce an
11
accurate yet real-time system in complex and dynamic environments such as public spaces. To cover large areas, the
12
AI-system would need to work with an integrated multi-camera setting, hence it must be robust enough to analyze video
or
13
14 streams that have varying characteristics due to the camera’s deployment environments.
15 People detection, tracking, and crowd counting originates from classifying and localizing people continuously
ma
16 across the captured image streams, commonly known as object detection (Hu et al., 2018). Current object detection
17 models can be categorized into three types: two-stage, one-stage and mobile-based models. Two-stage models are
18 known for their very accurate detection but slower inference time in applications (Ren et al., 2015, Zhou et al., 2017).
19
One-stage models are designed to improve the inference time of two-stage models while still achieving comparable
20
tio

21 accuracy (Liu et al., 2016, Redmon and Farhadi, 2018). Mobile-based models are the compact version of the previous
22 models dedicated to specific applications on smaller computing devices such as mobile and edge devices (Wang et al.,
23 2018). Mobile-based models are built on one-stage model with smaller neural networks for faster and cheaper
nT

24 computation (Biswas et al., 2019). People tracking applies the detection model into the image stream using specific
25 tracking algorithms to stitch the detection information. The key in people tracking is to lock onto every targeted object
26 in the video frame, uniquely identify each one of them and track them until they exit the video frame. Spatio-temporal
27
features are essential for this tracking process. Existing work can be categorized into traditional and deep learning
ec

28
29 methods.
30 Traditional methods encompass mean-shift, optical flow, and Kalman filter. Mean-shift algorithms examine the
hn

31 distribution mode of features of an object and track the change in the distribution across the frames to track the object
32 (Iswanto and Li, 2017). Optical flow, on the other hand, does not necessarily use the object features but the Spatio-
33 temporal brightness variations of the image frames (Yang et al., 2019b). It applies an equation and some prediction
34 techniques across the frames to track the object. Kalman filter is well known to solve temporal-related vision machine
olo

35
learning problems (Gunjal et al., 2018). It uses the available detections and prior predictions to predict the best possible
36
37
tracking, while considering the error probability of the task. Despite the usefulness of Kalman filter, it is unsuitable for
38 real-world applications due to occlusions and different viewpoints in the integrated surveillance system.
gy

39 Prominent deep learning methods for object tracking are deep regression networks, Recurrent YOLO, Simple
40 Real-time Tracker (SORT) and Deep SORT (Bewley et al., 2016, Held et al., 2016, Wojke et al., 2017, Yun and Kim,
41 2019). Deep regression networks are one of the early deep learning methods. It uses a two-frame Convolutional Neural
42 Network (CNN) architecture on current and previous video frames to regress and localize the targeted objects. Recurrent
&P

43 YOLO uses a one-stage detection model, YOLO, and combines it with LSTM to process both spatial and temporal
44
features of the images providing significant improvements on early tracking methods. SORT is designed to handle object
45
46 tracking in real-time applications, by leveraging the benefit of CNN for detection with Kalman filter and Hungarian
47 algorithm to improve tracking performance. The pragmatic SORT approach results in comparable accuracy to state-of-
eo

48 the-art methods and faster computation. Deep SORT is an improvement on SORT, where it incorporates deep
49 association metrics pre-training to improve the tracking accuracy during occlusions. The deep metric lowers the chance
50 of false tracking by ensuring the system tracks the correct person despite repeated occlusions.
ple

51 Social distancing requirements due to COVID-19 have led to some developments in vision-based technology.
52
An initial study compared sensor-based and vision-based systems on pinpointing coordinates of people using AI
53
54 (Elloumi et al., 2016, Johnson Jr et al., 2020). The vision-based system demonstrated superior accuracy, where it uses
55 images captured from smartphone cameras and applied people detection and Kalman Filter to compute the coordinates.
56 Another idea is to use drones to track people and measure their temperature to determine the likelihood of COVID-19
57 infection, but this is still under development and not designed to manage social distancing (Daly, 2020). Another
58 proposed approach was to track people’s movement from cameras, but privacy issues were not addressed as it used a
59 centralized computation approach (Punn et al., 2020).
60
2.2. Law and Ethics for Privacy
 Information Technology & People Page 8 of 44

Data privacy is still an open and challenging issue in AI, especially when dealing with sensitive and personal
1
data. The use of AI for business or by any organization has been deemed a threat to people’s privacy as a system’s
2
3
intelligence is reliant on gaining as much customer data as possible (Benson et al., 2015). For instance, many websites
4 have collected data regardless of customers’ consent by tracking online behavior and browsing history and sold it as a
5 commodity for marketing purposes (Liyanaarachchi, 2020). While wireless and sensor-based technologies require
6 consent from people before sharing their locations, vision-based technology collects data, including sensitive data such
7 as people’s faces, without explicit permission (Australian Government Office of the Australian Information
8 Commissioner, 2019). While collecting more data from cameras to update AI models can improve the models’
9 robustness, it also raises potential serious threats such as data misuse and data theft during transmission.
Inf
10
The increasing number of AI applications has led many countries, specifically Asia-Pacific countries and United
11
12 States of America (US), to establish privacy acts (Clifton et al., 2020). These regulations indicate the importance of
lawful data collection, processing, retention and use of personal data to protect people’s privacy. Fair and lawful data
or
13
14 collection is regulated across Asia-Pacific (Kennedys Law, 2018), the US (Department of Justice, 2015), and European
15 Union countries (General Data Protection Regulation, 2016), where they require people to know and understand when
ma
16 their data are collected. The acts indicate the importance of upholding data privacy for any collection and processing
17 system, as seen in the inclusion of privacy-by-design principle in the US Privacy Act of 1974. Any organizational body
18
needs to provide notice to people before or during data collection, or in certain countries such as Australia (Office of
19
20 the Australian Information Commissioner, 2019) and South Korea (Personal Information Protection Commission,
tio

21 2011), as soon as possible after collection. However, the latter case also requires de-identifying of personal data during
22 collection when possible. The majority of Asia-Pacific countries also require data de-identifying or deletion, specifically
23 after use. The privacy acts also encompass visual surveillance, particularly countries which surveillance systems are
nT

24 well established in public spaces, such as South Korea, Australia and Macau (Office for Personal Data Protection, 2005).
25 Therefore, it is critical and required by law to develop AI-enabled video surveillance systems without breaching data
26
privacy.
27
A recent study shows that data privacy can be protected using the federated learning approach (Yang et al.,
ec

28
29 2019a). The problem with the common machine learning approaches in the surveillance system is that data, such as
30 images, need to be sent to the central system’s storage and stored for a long time to augment the models’ knowledge.
hn

31 The application of intelligent vision application in public spaces will require the system to store and process countless
32 image streams in the central system storage. The federated learning approach enables AI to train an intelligent central
33 model from many smaller branch models by edge computing. This mechanism allows the central model to retrieve the
34
trained AI model (knowledge) from branch models and no longer store the actual image data without compromising
olo

35
36
model accuracy. Thus, any sensitive data in the image data will be immediately discarded from each camera and not
37 stored in the central system, upholding the law and ethics of data privacy.
38
gy

39 2.3. Summary of Current Gaps

40 Existing research on AI development to help maintain social distancing has not fully considered privacy
41 preservation. Removing person-identifiable information, such as facial data, and training the AI model on edge can
42 potentially support better privacy preservation. The proposed solution adopts on-edge computing to process video on
&P

43 the camera itself, so that no sensitive data is transmitted to the central machine. When processing is performed on edge,
44
the model is limited to learn from its local data and only benefits its device, hence the system still requires aggregation
45
46 of local model updates across edge devices as well. The federated learning approach enables aggregation of the local
47 model updates without sharing its actual data. For a large site like airports, aggregating local models from multiple
eo

48 cameras can improve the robustness of the overarching AI model.

49
50 3. Proposed Framework
ple

51 The proposed framework consists of three components, namely a design approach for responsible AI, a new AI
52 system designed for monitoring social distancing via video surveillance, and novel algorithms and architectures to
53
support the system.
54
55
56 3.1. Design Approach for Responsible AI-Enabled Video Surveillance
57 To promote technology integration, our system’s design and development was guided by a new Responsible AI
58 Implementation Framework (RAIIF – see Figure 1), which extends the human-centered design framework (Rosenbrock,
59 2012). The framework emphasizes iterative co-creation processes underpinned by trust establishment and maintenance.
60 It aims to streamline the AI adoption process by involving all stakeholders throughout the implementation process, from
planning and design to development and scaling-up. An iterative and agile approach applies to the whole journey of
Page 9 of 44 Information Technology & People

planning, design, development, deployment, operating, monitoring, and scaling of the AI solution. Each process may be
1
iterated on its own or after the whole journey (from planning to scaling) of the initial AI solution is completed.
2
3
The goal of the planning and designing process is to achieve AI trustability, which encompasses technological
4 accessibility and impacts consideration in the existing human and technology infrastructures. The initial AI models and
5 algorithms need to determine the business model’s viability, technological feasibility, and human desire (to adopt the
6 system). The development and deployment process aims to ensure that the AI solution is robust, consistently accurate,
7 meets expectations, and compliant until it reaches the point where the AI solution can be scaled up into full production.
8 The operating, monitoring, and scaling process aims to achieve optimized operations, empowered employees, engaged
9 customers, and transformed products. The objective is promoting AI solutions to harmoniously co-exist with businesses
Inf
10
and the community while maintaining individual and societal trust in the technology amid perceived risks.
11
12
---Insert Figure 1 here---
or
13
14
15 The framework provide checkpoints to observe the following responsible AI criteria.
ma
16 Reliable: the AI behaves as expected in every possible set of circumstances, even for novel inputs (unseen inputs
17 not covered in training data). When AI is used in public spaces, the AI solution must work well against any variations
18 such as occlusion, head pose, illumination, gender, or race. Low-reliable AI can result in misleading outcomes and
19 possible huge risks (e.g. loss of human life, finance loss, social risk) especially when deployed in critical and sensitive
20
tio

application domains such as in health. When the change is constantly introduced by the environment where AI is
21
22 deployed and significantly affects AI performance, continuous learning capability must be considered while developing
23 AI solutions (Parisi et al., 2019). In surveillance systems, continuous learning aims to allow the AI models to learn from
nT

24 new data collected from many cameras.

25 Non-biased: Biased inferences can affect AI’s reliability which can lead to unfairness issues such as racism
26 problems. When AI outcomes may be distorted, the AI users should be explicitly warned (Ghallab, 2019). Bias can
27 result from two sources: training data and algorithms. From the algorithm aspect, the nature of the algorithm used must
ec

28
be thoroughly investigated under any circumstances to ensure its reliability. Although deep learning has shown its high
29
30 accuracy in classification problems, it is a black box function approximator with limited interpretability as accuracy
drops to nearly zero in the presence of small adversarial perturbations (Raghunathan et al., 2018). From the data
hn

31
32 perspective, unbalanced training data distribution (e.g. age, gender, race) can cause the model to favor over-sampled
33 classes rather than under-sampled classes. In surveillance systems, collecting new data from cameras across different
34 locations can help to balance the dataset distribution issue.
olo

35 Data privacy protection: Privacy of personal data is essential when accessing sensitive and personal data, such
36
as facial data (Domingo-Ferrer and Blanco-Justicia, 2020). According to General Data Protection Regulation (GDPR),
37
38
data privacy breaches happen when collected data are not used for the targeted purposes (Sultan and Jensen, 2020) or
accessed by unauthorized people. Such protection must be taken as a proactive approach rather than a reactive approach,
gy

39
40 meaning it must be anticipated before any individuals are affected or hurt (Sultan and Jensen, 2020). Privacy protection
41 of sensitive data must be applied in any data activities, from collecting (when data are captured by surveillance cameras),
42 processing (when data are processed and analyzed at a processing machine), storing (when data are retained for period
&P

43 of time at a machine) to moving data activities (when data are transferred from cameras to a central or cloud machine).
44 Therefore, surveillance systems must take this responsibility into account when the system is initially designed until
45
fully deployed in operation. A poorly-design surveillance system portrays a false sense of security and violates the
46
47 fundamental privacy rights of individuals (General Data Protection Regulation, 2016).
eo

48 Transparent AI decisions: Transparency of how AI works in making decisions can help organizations be aware
49 of risks and mitigate issues of fairness and trust. Transparency of how AI models make automatic decisions means
50 understanding the assumptions, limitations, and criteria. The AI must be able to explain and justify the response
ple

51 generated to the input. Moreover, transparency is also required in the context of data privacy and security (Burt, 2019).
52 How the system deals with any identifiable and sensitive data must be carefully designed and any issues occuring must
53
be highlighted. This must be done in terms that are understandable to the users. More transparency can help reduce
54
55 people’s unfounded fears of AI (Ghallab, 2019). By knowing this, the organization can decide how far the AI system
56 will do the automation. In some cases, a human-in-the-loop approach may be required during the process of decision
57 making as AI progresses until humans cannot add new value to it (Ransbotham, 2017).
58 Sustainable and scalable for long-term use: Sustainable and scalable AI can contribute to developing trust for
59 users. High investment in AI requires high commitment from stakeholders. Therefore, people demand high assurance
60 that the technology can be scaled up to contribute in wider ranges and longer periods of time. The acceptability of a
technology needs to take into account the long term and global constraints (Ghallab, 2019). The technology
infrastructure can be used across different levels of organizations.
 Information Technology & People Page 10 of 44

Table II to Table IV outline the checklists for each process to enforce these criteria.
1
2
---Insert Table II to Table IV here---
3
4
5
3.2. System Design
6 Figure 2 illustrates how the system preserves privacy by ensuring that each camera processes video on edge
7 device, known as on-edge computing, and only uses representative data for calculating social distancing.
8
9 ---Insert Figure 2 here---
Inf
10
11 The purpose of on-edge computing is to enable surveillance cameras with computation capability. It aims to
12 allow any computations to be performed at the device hence eliminating the dependency of cloud computing at the
or
13 central server. Such approaches preserve any sensitive and identifiable data to reside at the device where they are
14
collected. There are two computations performed on edge devices: 1) to detect social distancing breaches and 2) to
15
perform AI continuous learning.
ma
16
17 To meet the AI reliability criterion, the system employs AI continuous learning to continuously adapt the people
18 detector model against any changes introduced by the environment. New images are periodically collected from cameras
19 across locations to build new training data that represent any variation and to meet the non-biased inference criterion.
20
tio

To allow continuous learning without transmitting any sensitive and identifiable data, a federated learning approach
21 (Konečný et al., 2016) is employed to meet the data privacy protection criterion.
22
Federated learning enables smart cameras to continuously update their local models by performing local
23
inference and training within the device and update the central model by aggregating local model updates without
nT

24
25 sending actual images. An edge device aims to collect and process data from a camera and perform local training to
26 update its local model. The central machine has one global model and aims to aggregate knowledge from edge devices’
27 local models. The central machine does not have direct access to private data from surveillance cameras.
ec

28 Figure 3 explains the process of detecting social distancing breaches within each smart camera. Initially,
29 people’s locations are detected within frames by a pedestrian detector model (Step 1). For any people detected, each
30
person is tracked over frames using an object tracking model (Step 2) and an anchor point is determined based on the
hn

31
32 person’s foot position representing the person’s position in 2D projection (Step 3). Any anchor points are then
33 transformed to 2D projection using matrix H (Step 4). Figure 4 illustrates how the camera scene is projected from 3D
34 to 2D scene using the Homography algorithm. The number of people is then calculated based on people being tracked
olo

35 (Step 5) and crowdedness level is calculated (Step 6). Spatial distance between people is then calculated using Euclidean
36 distance (Step 7). Based on the distance, social distance breaches can be determined (Step 8). Any people that are below
37 the safe distance threshold will be drawn in green. Any people that are higher than the safe distance will be drawn in
38
red including the distance. To provide social distancing breach analysis, a heat map is generated based on cumulative
gy

39
40
distancing breach detection and updated periodically (Step 9). A breaching score is calculated where the incident
41 happens (between 0-1). The closer the distance is, the higher the breaching score. The breaching score will be
42 accumulated throughout the day and will be reset to 0 for the next day. Five levels of breach are used to visualize the
&P

43 area, from low to critical breach level. Breach level is determined by the number of distancing breach occurrences and
44 how close the breaches are. The closer the distancing breach, the higher the multiplier put into the calculation. All spatial
45 information is stored in a JSON file and then transmitted to the server (Step 10). When the breach becomes critical,
46 automatic notifications will be triggered and sent to authorized people via email and phone (Step 11).
47
eo

48
---Insert Figures 3 and 4 here---
49
50
ple

51 3.3. On-edge Algorithms and System Architecture

52 The system employs two cutting-edge algorithms to detect people (Step 1 at Figure 3) and track people (Step 2
53 at Figure 3) to achieve robust performance. Lightweight computation without compromising accuracy is the main
54 consideration while employing algorithms for on-edge deployment.
55 People detector model: SSD-MobileNet V2 object detector model is employed to detect people within frames
56 (Sandler et al., 2018). The model is a single-state object detector which detects 90 types of objects from image pixels
57
and returns the detected objects’ bounding box coordinates and class confidence. It was trained on MS-COCO dataset,
58
59 a large-scale object dataset consisting of 300,000 images. For our purpose, we only used the person (or pedestrian) class
60 and disregarded the other classes. The model architecture is based on inverted residual structure where input and output
of residual blocks are thin bottleneck layers as opposed to traditional residual models. It refines its initial model by
replacing expensive-to-compute convolutional layers with depth-wise separable convolution blocks. Compared to
Page 11 of 44 Information Technology & People

similar networks, the model has about 9 times less computation while achieving same accuracy with higher FPS,
1
enabling real-time processing.
2
3
People tracking model: Simple Online and Realtime Tracking with a deep association metric (Deep SORT)
4 model is employed to track people over frames (Wojke et al., 2017). The purpose of tracking people is to count the same
5 person as one object during its movement over frames. It is an online object tracking approach which is focused on
6 simple yet effective real-time tracking. Improved from its initial version, the algorithm integrates visual appearance
7 information aiming to improve performance by tackling longer periods of occlusion in the real world yet reducing the
8 number of identity switches by 45%.
9 Figure 5 outlines important steps in how to initiate and perform federated learning. The steps are repeated
Inf
10
periodically to keep updating the models based on newly collected data.
11
12
---Insert Figure 5 here---
or
13
14
15 Step 1: At the central machine, a base model with pre-trained weights is adopted as the initial global model. The
ma
16 model is initially trained and fine-tuned using existing people datasets. Then, the global model is deployed to all edge
17 devices as their initial model. This is only executed once. Whenever a new camera is deployed, the most up-to-date
18 global model can be deployed as the initial local model.
19 Step 2: At each edge device, continuous training is performed periodically (managed by a scheduler) using a
20
tio

fine tuning approach (Käding et al., 2016) to improve the local model’s performance by learning from newly collected
21
22 data. The model update will be then generated once local training is finished and all local images are immediately
23 removed.
nT

24 Step 3: The central machine will periodically notify all edge devices if they have any local updates and request
25 their participation in sending local updates to update the global model. Edge devices may refuse to participate when
26 they do not have any latest local model updates or for any other reasons.
27 Step 4: For any edge devices that are willing to participate, they respond to the central machine and send their
ec

28
local model’s update obtained from local training. Any data compression and encryption methods can be applied to
29
30 secure data sent during transmission and avoid data misuse.
Step 5: The central machine updates its global model by aggregating all local model updates using an
hn

31
32 aggregation algorithm (Wang et al., 2020, Abadi et al., 2016). Once this step is finished, it will return to Step 2 in the
33 next cycle
34 Edge devices can schedule local training when the device is idle. Ideally, night is best time to perform local
olo

35 training but this depends on the context and the business operating hours. Schedules can be set manually according to
36
each condition but only executed when in idle operation.
37
38
4. Results and Evaluation
gy

39
40 The proposed system is evaluated in terms of 1) its reliability by validating that the algorithm can achieve state-
41 of-the-art performance in the context of the case study, 2) its usefulness based on its robustness for real deployment in
42 the airport’s selected cameras, and 3) its compliance with responsible AI implementation, based on meeting criteria and
&P

43 checklists. It should be noted that the results are presented with people’s faces are blurred to protect their identities in
44 accordance with the ethics agreement.
45
46 Automatic people detection and tracking
47
eo

Table V shows the performance comparison of object tracking algorithms evaluated on the Multiple Object
48
49 Tracking Benchmark 2016 (MOT16 challenge). Several metrics are used to evaluate the model performance. Compared
50 to the initial version, DeepSort could reduce identity switches (ID) by 45%, from 1423 to 781, and increase track
ple

51 fragmentation (FM) slightly. There is also significant increase in the number of mostly tracked objects (MT) and a
52 decrease in mostly lost objects (ML). The incorporation of visual appearance information could maintain identities
53 longer against the occlusion in real-world settings. In terms of run time, half the time was spent on feature generation.
54 However, the use of GPU embedded in the edge device will significantly speed up computation hence achieve real-time
55
processing. A further detailed comparison can be found in the referred paper (Wojke et al., 2017).
56
57
58 ---Insert Table V here---
59
60 The performance was evaluated on four cameras by observing how people are detected and tracked by the
system compared to ground truth. We were unable to evaluate the automatic people detection and tracking since there
is no available annotation of the airport’s cameras. Therefore, performance evaluation will be further investigated in
automatic crowd counting and social distancing breach detection (discussed in the next two sections) since people
 Information Technology & People Page 12 of 44

detection and tracking performance will greatly affect crowd counting and social distancing breach detection
1
performance.
2
3
Through observation and experiments, we found that camera angle plays as an important role in obtaining good
4 people detection. About 45 to 60 degree is a good camera angle to capture, detect and track people’s movement in a
5 public area. Detection accuracy dropped significantly when tested on side-placed cameras or too-close cameras
6 especially in busy times (e.g. departing time at departure gate) as seen in Figure 6. That number of people will
7 immediately create high numbers of occlusions; hence cameras cannot detect people independently. Therefore, detection
8 accuracy can be maintained when the camera is placed far enough away with good angle (bird-eye) to reduce occlusion.
9
Inf
10 ---Insert Figure 6 here---
11
12
Automatic crowd counting
or
13
14 The performance was evaluated on two cameras by validating the number of people counted by the system
15 compared to ground truth. Several different algorithms were used for comparison. For detecting people, three pedestrian
ma
16 detector models (YoloV3, SSD300, and MobileV2 SSD) are employed. For tracking people, three object tracker models
17 (KCF, SORT, and Deep SORT) are employed. The evaluation was tested on 30-minute video feed from each camera in
18 a standard busy hour. For ground truth data, the number of people is counted by manual observation. Table VI presents
19 the accuracy on two cameras using four combinations of algorithms. Figure 7 shows some samples of crowd counting
20
tio

in several cameras.
21
22
23 ---Insert Table VI here---
nT

24
25 ---Insert Figure 7 here---
26
27 When tested on CPU, YoloV3 + DeepSort has the best accuracy but requires the most expensive computation
ec

28 (nearly 60 min computation time for processing all frames). YoloV3 + Sort has good accuracy but cannot have real-
29 time computation as it requires less time (nearly 55 min) to process all frames. SSD300 + KCF has fastest computation
30 time (up to 10 min) but has the worst accuracy (Figure 8).
hn

31
32
---Insert Figure 8 here---
33
34
olo

35 When tested on Google Coral Dev Board, MobileV2 + DeepSort can detect and track people on all frames
36 achieving real-time computation. It was tested on videos with dimensions of 1,920 x 1,080 pixels and 12 fps. Pre-
37 processing is performed to reduce unnecessary computation. Video is downsized to about half size where people objects
38 are still identifiable and detection accuracy is not affected. As a result, the device can achieve 9 fps, enabling real-time
gy

39 computation on edge devices. In addition, the model is initially converted to TPU model so inference can work much
40 faster compared to CPU computation.
41
42
Automatic social distancing breach detection
&P

43
44
Performance was evaluated in busy areas by validating social distancing breach detection by the system
45 compared to ground truth. Three cameras were used for evaluation: check-in area, food court area, and waiting area
46 where people may gather in large numbers. Since it is difficult to annotate ground truth computationally, manual
47 observation was performed by humans through video feeds to ensure if social distancing breaches occurred. Two people
eo

48 were employed to carefully observe and compare the actual video and analysis results to determine if people marked as
49 red were breaching safe distances alongside the heat maps and automatic notifications generated. Figure 9 shows
50 samples of social distancing breach detection including heat maps at a particular time on check-in and food courts
ple

51
cameras, respectively. Figure 10 shows how people detection model detects walking and sitting people on the departure
52
53 waiting area camera. It confirms the importance of continuous learning capability in the proposed system to periodically
54 re-adapt the AI models.
55
56 ---Insert Figures 9 and 10 here---
57
58 Checklist for meeting the criteria for Responsible AI
59 The proposed technology was developed while involving the stakeholders that can help to ensure responsibility
60 towards building trustworthy AI.
Reliable: Federated learning supports continuous learning aiming to improve AI reliability by learning from
periodically collected new data without compromising data privacy and model accuracy.
Page 13 of 44 Information Technology & People

Non-biased: Through federated learning, data bias can be gradually reduced by collecting more data for under-
1
sample classes while respecting data privacy. Human involvement (e.g. experts in observing surveillance) can reduce
2
3
AI bias by periodically evaluating the system’s output. Automatic notification heat maps also can help the airport to
4 evaluate bias e.g. manual in-field observation/checking to double check. Continuous benchmarking in different datasets
5 can help understand and measure fairness (e.g. evaluate the model using many different types of camera, different
6 datasets). Any previously uncaptured bias/unintentional bias can be discussed through regular meetings so airport staff
7 can understand and be aware of the biases. Involving third parties can help to determine if the algorithm works as
8 expected (e.g. publishing peer-reviewed papers, presenting at scientific conferences).
9 Data privacy protection: Through federated learning, privacy and use of personal data are managed through all
Inf
10
data activities. From collecting, processing, storing activities, all sensitive data and personal data are kept secure in local
11
12 devices where the data are obtained. For moving data, only non-sensitive data are sent to the central data, not actual
images. Data compression and aggregation can be applied to ensure data are more secure. Local training requires shorter
or
13
14 training time and limited movement of sensitive data. Hence, it shortens the duration of data retention as images are
15 immediately erased after training and minimizes the risks of data treats.
ma
16 Transparent AI decisions: Human-in-the-loop approach (i.e. integrating airport staff and the AI system to work
17 together in making decisions) will allow humans to double check results or make decisions. For example, auto
18
notification can ask airport personnel to validate how the system works. Working closely with airport management
19
20 through regular meetings helps build transparency in AI models.
tio

21 Sustainable and scalable for long-term use: The incorporation of AI continual learning capability enables
22 system to scale up in the future such as adding more cameras with verified camera deployment characteristics.
23 Furthermore, the system is not limited to monitoring social distancing during the COVID-19 pandemic, but can be
nT

24 adjusted for other purposes such as to observe people’s social interactivity at home or in the workplace.
25 Table VII to Table IX demonstrate how trust is established and maintained throughout the process by following
26
the checklists provided in the framework.
27
ec

28
29 ---Insert Tables VII, VIII, and IX here---
30
5. Conclusion and Future Work
hn

31
32 This paper proposes on-edge privacy-preserving AI-enabled surveillance technology to detect social distancing
33 breaches in public spaces to respect people’s privacy. Comprehensive evaluation in the context of a case study at an
34 airport demonstrates how the proposed system can fully address the case study’s requirements in terms of its reliability,
olo

35 usefulness when deployed to the airport’s cameras, and compliance with responsible AI. The results confirm that
36
limiting sensitive data movement via on-edge computing is effective to protect data privacy without compromising the
37
38
performance. This paper also contributes to the study on addressing data privacy issue in AI-enabled surveillance
applications. Existing work on automated social distancing monitoring has not fully considered the need for addressing
gy

39
40 privacy-preserving AI. While the proposed system has shown the benefits for monitoring social distancing during
41 pandemics, it can be adjusted for other purposes in the future, such as observing people’s social interactivity at home or
42 in workplaces that respects people’s privacy. The proposed system and its design approach can be also useful for
&P

43 informing future work in developing AI-enabled video surveillance applications, in which data privacy can be protected
44 using the federated learning approach, while ensuring its compliance with responsible AI principles.
45
46
References
47
eo

Al-Suwaidi, G. B. & Zemerly, M. J. Locating friends and family using mobile phones with global positioning system (GPS). 2009 IEEE/ACS
48 International Conference on Computer Systems and Applications, 2009. IEEE, 555-558.
49 Australian Government Office of the Australian Information Commissioner. 2019. Security cameras [Online]. Australian Government Office of
50 the Australian Information Commissioner. Available: https://www.oaic.gov.au/privacy/your-privacy-rights/surveillance-and-
ple

51 monitoring/security-cameras/ [Accessed].
52 Benson, V., Saridakis, G. & Tennakoon, H. 2015. Information disclosure of social media users. Information Technology & People.
53 Bewley, A., Ge, Z., Ott, L., Ramos, F. & Upcroft, B. Simple online and realtime tracking. Image Processing (ICIP), 2016 IEEE International
Conference on, 2016. IEEE, 3464-3468.
54 Biswas, D., Su, H., Wang, C., Stevanovic, A. & Wang, W. 2019. An automatic traffic density estimation using Single Shot Detection (SSD) and
55 MobileNet-SSD. Physics and Chemistry of the Earth, Parts A/B/C, 110, 176-184.
56 Boland, B., De Smet, A., Palter, R. & Sanghvi, A. 2020. Reimagining the office and work life after COVID-19.
57 Burt, A. 2019. The AI transparency paradox. Harv. Bus. Rev.(Dec. 13, 2019), https://bit. ly/369LKvq.
58 Clifton, J., Glasmeier, A. & Gray, M. 2020. When machines think for us: the consequences for work and place. Oxford University Press UK.
59 Daly, N. 2020. A 'pandemic drone' and other technology could help limit the spread of coronavirus and ease restrictions sooner, but at what cost?
[Online]. ABC News. Available: https://www.abc.net.au/news/2020-05-01/new-surveillance-technology-could-beat-coronavirus-but-
60
at-a-cost/12201552 [Accessed 26 Oct 2020 2020].
Department of Health, A. G. 2020. Physical distancing for coronavirus (COVID-19) [Online]. Australia Government Department of Health.
Available: https://www.health.gov.au/news/health-alerts/novel-coronavirus-2019-ncov-health-alert/how-to-protect-yourself-and-
others-from-coronavirus-covid-19/physical-distancing-for-coronavirus-covid-19 [Accessed].
 Information Technology & People Page 14 of 44

Department of Justice, T. U. S. 2015. Privacy Act of 1974. Department of Justice of The United States
1 Domingo-Ferrer, J. & Blanco-Justicia, A. 2020. Privacy-Preserving Technologies. The Ethics of Cybersecurity. Springer, Cham.
2 Elloumi, W., Latoui, A., Canals, R., Chetouani, A. & Treuillet, S. 2016. Indoor pedestrian localization with a smartphone: A comparison of inertial
3 and vision-based methods. IEEE Sensors Journal, 16, 5376-5388.
Feldstein, S. 2019. The global expansion of AI surveillance, Carnegie Endowment for International Peace Washington, DC.
4
General Data Protection Regulation. 2016. REGULATION (EU) 2016/679 OF THE EUROPEAN PARLIAMENT AND OF THE COUNCIL of 27
5 April 2016 [Online]. Official Journal of the European Union. Available: https://eur-lex.europa.eu/legal-
6 content/EN/TXT/PDF/?uri=CELEX:32016R0679 [Accessed].
7 Ghallab, M. 2019. Responsible AI: requirements and challenges. AI Perspectives, 1, 3.
8 Gunjal, P. R., Gunjal, B. R., Shinde, H. A., Vanam, S. M. & Aher, S. S. Moving Object Tracking Using Kalman Filter. 2018 International
9 Conference On Advances in Communication and Computing Technology (ICACCT), 2018. IEEE, 544-547.
Inf
10 Held, D., Thrun, S. & Savarese, S. Learning to track at 100 fps with deep regression networks. European Conference on Computer Vision, 2016.
Springer, 749-765.
11 Hu, H., Gu, J., Zhang, Z., Dai, J. & Wei, Y. Relation networks for object detection. Proceedings of the IEEE Conference on Computer Vision and
12 Pattern Recognition, 2018. 3588-3597.
or
13 Iswanto, I. A. & Li, B. 2017. Visual object tracking based on mean-shift and particle-Kalman filter. Procedia computer science, 116, 587-595.
14 Johnson Jr, J., Hasan, S., Lee, D., Hluchan, C. & Ahmed, N. 2020. Social-Distancing Monitoring Using Portable Electronic Devices.
15 Jovicic, A. & Engineer, P. Qualcomm ® Lumicast TM : A high accuracy indoor positioning system based on visible light communication April
2016. 2016.
ma
16
Kavalionak, H., Gennaro, C., Amato, G., Vairo, C., Perciante, C., Meghini, C. & Falchi, F. 2019. Distributed video surveillance using smart
17 cameras. Journal of Grid Computing, 17, 59-77.
18 Kennedys Law. 2018. Summary of Data Privacy Laws in Asia Pacific [Online]. Available: https://kennedyslaw.com/media/3154/asia-pacific-data-
19 principles.pdf [Accessed 26 Oct 2020 2020].
20
tio

Kok, M., Hol, J. D. & Schön, T. B. 2015. Indoor positioning using ultrawideband and inertial measurements. IEEE Transactions on Vehicular
21 Technology, 64, 1293-1303.
22 Konečný, J., Mcmahan, H. B., Ramage, D. & Richtárik, P. 2016. Federated optimization: Distributed machine learning for on-device intelligence.
arXiv preprint arXiv:1610.02527.
23
Laoudias, C., Moreira, A., Kim, S., Lee, S., Wirola, L. & Fischione, C. 2018. A survey of enabling technologies for network localization, tracking,
nT

24 and navigation. IEEE Communications Surveys & Tutorials, 20, 3607-3644.

25 Ledergerber, A. & D’andrea, R. 2020. A Multi-Static Radar Network with Ultra-Wideband Radio-Equipped Devices. Sensors, 20, 1599.
26 Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu, C.-Y. & Berg, A. C. Ssd: Single shot multibox detector. European conference on
27 computer vision, 2016. Springer, 21-37.
ec

28 Liyanaarachchi, G. 2020. Online privacy as an integral component of strategy: allaying customer fears and building loyalty. Journal of Business
Strategy.
29
Mazuelas, S., Bahillo, A., Lorenzo, R. M., Fernandez, P., Lago, F. A., Garcia, E., Blas, J. & Abril, E. J. 2009. Robust indoor positioning provided
30 by real-time RSSI values in unmodified WLAN networks. IEEE Journal of selected topics in signal processing, 3, 821-831.
hn

31 Nadikattu, R. R., Mohammad, S. M. & Whig, D. 2020. Novel Economical Social Distancing Smart Device for COVID19. International Journal
32 of Electrical Engineering and Technology, 11.
33 Niu, J., Wang, B., Shu, L., Duong, T. Q. & Chen, Y. 2015. ZIL: An energy-efficient indoor localization system using ZigBee radio to detect WiFi
34 fingerprints. IEEE Journal on Selected Areas in Communications, 33, 1431-1442.
olo

35 Office for Personal Data Protection 2005. Personal Data Protection Act. In: REGION, O. F. P. D. P. O. T. M. S. A. (ed.). Office for Personal Data
Protection.
36 Office of the Australian Information Commissioner 2019. Australian Privacy Principle Guidelines. In: COMMISSIONER, O. O. T. A. I. (ed.).
37 Office of the Australian Information Commissioner.
38 Parisi, G. I., Kemker, R., Part, J. L., Kanan, C. & Wermter, S. 2019. Continual lifelong learning with neural networks: A review. Neural Networks,
gy

39 113, 54-71.
40 Personal Information Protection Commission 2011. Personal Information Protection Act. Ministy of Government Legislation of Korea.
41 Punn, N. S., Sonbhadra, S. K. & Agarwal, S. 2020. Monitoring COVID-19 social distancing with person detection and tracking via fine-tuned
YOLO v3 and Deepsort techniques. arXiv preprint arXiv:2005.01385.
42 Raghunathan, A., Steinhardt, J. & Liang, P. 2018. Certified defenses against adversarial examples. arXiv preprint arXiv:1801.09344.
&P

43 Ransbotham, S. 2017. Justifying Human Involvement in the AI Decision-Making Loop [Online]. Available:
44 https://sloanreview.mit.edu/article/justifying-human-involvement-in-the-ai-decision-making-loop/ [Accessed].
45 Redmon, J. & Farhadi, A. 2018. Yolov3: An incremental improvement. arXiv preprint arXiv:1804.02767.
46 Ren, S., He, K., Girshick, R. & Sun, J. Faster r-cnn: Towards real-time object detection with region proposal networks. Advances in neural
47 information processing systems, 2015. 91-99.
eo

Romm, T., Harwell, D., Dwoskin, E. & Timberg, C. 2020. Apple, Google debut major effort to help people track if they’ve come in contact with
48
coronavirus [Online]. The Washington Post. [Accessed 27 July 2020 2020].
49 Rosenbrock, H. H. 2012. Designing human-centred technology: a cross-disciplinary project in computer-aided manufacturing, Springer Science
50 & Business Media.
ple

51 Rossi, F. 2018. Building trust in artificial intelligence. Journal of international affairs, 72, 127-134.
52 Sanchez-Iborra, R. & Skarmeta, A. F. 2020. TinyML-Enabled Frugal Smart Objects: Challenges and Opportunities. IEEE Circuits and Systems
53 Magazine, 20, 4-18.
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A. & Chen, L.-C. Mobilenetv2: Inverted residuals and linear bottlenecks. Proceedings of the IEEE
54
conference on computer vision and pattern recognition, 2018. 4510-4520.
55 Sugianto, N., Tjondronegoro, D., Sorwar, G., Chakraborty, P. & Yuwono, E. I. 2019. Continuous Learning without Forgetting for Person Re-
56 Identification. In: ABADI, M. (ed.) 16th IEEE International Conference on Advanced Video and Signal-based Surveillance. Taipei,
57 Taiwan: IEEE.
58 Sultan, S. & Jensen, C. D. Privacy-preserving measures in smart city video surveillance systems. 6th International Conference on Information
59 Systems Security and Privacy, 2020. SciTePress, 506-514.
60 Szajewska, A. 2017. Development of the Thermal Imaging Camera (TIC) Technology. Procedia Engineering, 172, 1067-1072.
Tiemann, J. & Wietfeld, C. Scalable and precise multi-UAV indoor navigation using TDOA-based UWB localization. 2017 International
Conference on Indoor Positioning and Indoor Navigation (IPIN), 2017. IEEE, 1-7.
Wang, R. J., Li, X. & Ling, C. X. Pelee: A real-time object detection system on mobile devices. Advances in Neural Information Processing
Systems, 2018. 1963-1972.
Page 15 of 44 Information Technology & People

World Health Organization. 2020. Coronavirus disease (COVID-19) pandemic [Online]. Available:
1 https://www.who.int/emergencies/diseases/novel-coronavirus-2019 [Accessed 19 Oct 2020].
2 Wojke, N., Bewley, A. & Paulus, D. Simple online and realtime tracking with a deep association metric. Image Processing (ICIP), 2017 IEEE
3 International Conference on, 2017. IEEE, 3645-3649.
Yang, Q., Liu, Y., Chen, T. & Tong, Y. 2019a. Federated machine learning: Concept and applications. ACM Transactions on Intelligent Systems
4
and Technology (TIST), 10, 12.
5 Yang, T., Cappelle, C., Ruichek, Y. & El Bagdouri, M. 2019b. Online multi-object tracking combining optical flow and compressive tracking in
6 Markov decision process. Journal of Visual Communication and Image Representation, 58, 178-186.
7 Yun, S. & Kim, S. Recurrent YOLO and LSTM-based IR single pedestrian tracking. 2019 19th International Conference on Control, Automation
8 and Systems (ICCAS), 2019. IEEE, 94-96.
9 Zhou, X., Gong, W., Fu, W. & Du, F. Application of deep learning in object detection. 2017 IEEE/ACIS 16th International Conference on
Inf
10 Computer and Information Science (ICIS), 2017. IEEE, 631-634.
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 16 of 44

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23 Figure 1. Responsible AI Implementation Framework (RAIIF). The framework is iterative across all processes from
nT

24 planning and designing to development and scaling-up processes.

25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 17 of 44 Information Technology & People

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25 Figure 2. The proposed system for detecting social distancing breach in public space. Left: People detected on
26 cameras. Right: People detected are represented as dots in a 2D projection view to make them feel less observed.
27 Green identifies a person is at a safe distance whereas red identifies a person at an unsafe distance.
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 18 of 44

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15
ma
16
17
18 Figure 3. The process of detecting social distancing breaches within each smart camera.
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 19 of 44 Information Technology & People

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
Figure 4. Area projection alignment from 3D to 2D projection using the Homography (H) algorithm. During the
14
15
initial camera setup, corner points (red dots) are defined to calculate the matrix H.
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 20 of 44

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37 Figure 5. On-edge computing to preserve data privacy
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 21 of 44 Information Technology & People

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15 Figure 6. With side-place cameras or too-close cameras, people detection accuracy can be maintained when few
people are present at one time (left) but will drop significantly with too many people as occlusion happens (right).
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 22 of 44

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15 Figure 7. Samples of people detection and tracking by airport cameras. Left: arrival gate camera. Right: after-security
check camera
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 23 of 44 Information Technology & People

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15 Figure 8. Accuracy (left) and computation time (right) comparison of people counting in arrival gate camera.
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 24 of 44

1
2
3
4
5
6
7
8
9
Inf
10
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
Figure 9. Sample of social distancing breach detection by two cameras. Top: check-in area camera. Bottom: café at
23
departure camera. Images are captured at a particular time and the heat map shows cumulative breaches on that day.
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 25 of 44 Information Technology & People

1
2
3
4
5
6
7
8
9
Inf
10
11
12 Figure 10. Sample of social distancing breach detection by the departure waiting area camera. Images are captured at
or
13 a particular time and the heat map shows cumulative breaches on that day.
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 26 of 44

Table I. Evolution of AI solutions adopted at the local international airport.

1
2 Business requirement Proposed AI Benefit Consideration
3 solution
4
5 People are detected Automatic people The airport can observe passenger Accuracy of people detection
6 accurately by different detection and flow (e.g. queuing duration) and tracking
7 cameras and deployments tracking
8
9 Crowd counting is accurate Automatic crowd The airport can monitor the level of Accuracy of crowd counting
Inf
10 for different cameras and counting crowdedness in each area to maintain
11 deployments safe distancing.
12
Social distancing breach is Automatic social The airport can enforce social Accuracy of the social
or
13
detected accurately based on distancing breach distancing responsibility and distancing breach detection and
14
government regulations detection rethink/redesign the space layout to spatial analysis
15
during COVID-19 pandemic reduce distancing breaches
ma
16
17 (e.g. < 1.5 m)
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 27 of 44 Information Technology & People

Table II. Checklist for Plan and Design stages

1
2 Consideration Checklist
3
4 Technological accessibility  The organization's existing internal infrastructure can support the AI solution, including people
5 and data.
6  The organization commits to transform its operational processes to align with the AI solution
7 and its required technologies, including data collection and management.
8
9 AI-enabled business  The initial solution can confirm the AI’s feasibility and ability to serve a purpose, such as
Inf
10 refinement solving an identified problem or supporting business processes.
11  The organization's decision-makers understand the business case of the AI solution, in terms
12 of its potential for delivering a return of investment, such as cost-saving, enhanced decisions,
or
13 and improved business processes' efficiency.
14  Decision-makers understand the potential impacts on the existing human and technology
15 infrastructure, and risks to the organization's reputation.
ma
16
17 Customer satisfaction  The AI solution brings direct benefits to end-users and improves customer experience.
18  Users can trust the AI solution and the organization that manages it due to clear manuals,
19 documentations, and transparent risk assessments.
20 
tio

Users can use the AI-enabled services, including the time for learning (how to use), and buy
21 the required equipment and fees to access it.
22
23 Operational compliance  AI solution is compliant with regulations and policies and operates lawfully.
nT

24  The impacts of operating and maintaining the AI solution can be managed by and compatible
25 with the existing human and technology infrastructure.
26  AI solution aims to augment human abilities to support actions and decisions while ensuring
27 the end-users' safety and wellbeing.
ec

28  AI solution is equitable and particularly accessible for those who need the most.
29
 AI solutions can be potentially used on a global scale without a significant shift in governance.
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 28 of 44

Table III. Checklist for Development and Deployment stages

1
2 Consideration Checklist
3
4 Reliable operations  AI decisions and reasoning can be ascertained and trusted to support operation, and bias can
5 be managed and reduced.
6  AI solution will function as intended and cannot be hacked and manipulated.
7
8 Empowered employees  AI solution can improve a human's ability to perform operational tasks and make decisions.
9  AI-based decisions are consistent, reliable, fair, and aligned with social and cultural justice
Inf
10 values in the local and global contexts.
11  If applicable, the change management for transforming the workforce always ensures fair
12 employment and labor practices.
or
13
14 Engaged customers  AI solution consistently meets performance expectations and increasingly more reliable to
15 support or reduce manual work, make better decisions, or make services more enjoyable and
ma
16 personalized.
17
18 Transformed products  AI solutions can transform the existing norms via the new product(s) that people can enjoy
19 and benefit from.
20  AI solution (potentially) adheres with the existing ethical- and legal- theories or general
tio

21 principles.
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 29 of 44 Information Technology & People

Table IV. Checklist for Operate, Monitor, and Scale-Up stages

1
2 Consideration Checklist
3
4 Continuous refinement  AI solution can adapt and refine its model, improving its accuracy based on human feedback.
5  AI solution can be continuously tested by research and development (R&D) and quality
6 assurance process to ensure that the outcomes continue to meet the required tasks and
7 performance expectations.
8  AI solutions can continuously adapt to new operations and socio-economic requirements.
9
Inf
10 Responsible business  AI solution can be managed by an end-to-end enterprise governance framework to be
11 practice consistently accountable and transparent.
12  AI's risks and scope of controls align with the organization's current and newly
or
13 operationalized ethics.
14
15 Individual (user) trust  Human trust cannot be broken by being disadvantaged, such as being subjected to automated
ma
16 decisions, especially when such decisions have legal ramifications.
17  Users have the right to access all information needed to fully understand the AI’s product and
18 the test results before adoption.
19  The policy and standard of practice are documented and transparent to ensure that users can
20
tio

trust the use of data and AI.

21
22 Societal (community) trust  An appropriate level of human supervision is maintained to ensure that the AI solution
23 continually aligns with human rights, social norms, and privacy regulations.
nT

24  Individual and collective AI use is transparent and auditable to ensure compliance with laws
25 and fundamental rights and freedom.
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 30 of 44

Table V. Performance comparison for object tracking algorithms evaluated on MOT16 challenge.
1
2 Algorithm MOTA ↑ MOTP ↑ IDF1 ↑ MT ↑ ML ↓ ID ↓ FM ↓ FP ↓ FN ↓ Runtime ↑
3
KCF (Chu et al., 2019) 48.8 - 47.2 - - 906 - 5875 86567 -
4
5 SORT (Bewley et al., 2016) 59.8 79.6 - 25.4% 22.7% 1423 1835 8698 63245 60Hz
6
7 Deep Sort (Wojke et al., 2017) 61.4 79.1 62.2 32.8% 18.2% 781 2008 12852 56668 40Hz
8
MOTA = Multi-object tracking accuracy, MOTP = Multi-object tracking precision, IDF1 = ID F1 score, MT = Mostly tracked object, ML = Mostly
9
lost object, ID = Identity switches, FM = Fragmentation. ↑ indicates the higher it is, the better it is whereas ↓ indicates the lower it is, the better it
Inf
10 is.
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 31 of 44 Information Technology & People

Table VI. People detection and tracking performance in the context of counting people in and out. Tested on 30-min
1
videos on both cameras.
2
3 Camera YoloV3 + DeepSORT YoloV3 + SORT SSD300 + KCF SSD MobileV2 + DeepSORT
4
5 Camera 1 70.55% 61.11% 53.91% 58.28%
6
(Arrival gate)
7
8 Camera 2 85.02% 80.05% 71.40% 74.53%
9
Inf
10 (After security check)
11
12
or
13
14
15
ma
16
17
18
19
20
tio

21
22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 32 of 44

Table VII. List of actions taken to establish and maintain trust in the Plan and Design stages
1
2 Checklist for Plan and Actions/Strategies taken
3 Design
4
5 Technological accessibility  Ensured existing IT infrastructure supports the proposed system, e.g. number of cameras
6 installed is sufficient to cover targeted observation area, number of personnel to observe,
7 hardware to process computation
8  Required the airport to commit to providing cameras’ access to develop the proof of concept
9 of the proposed AI solution
Inf
10
11 AI-enabled business  Provided a proof of concept to show the benefits of the proposed AI solution
12 refinement  Collaborated with the airport management through regular progress meetings and identify
or
13 and disclose any impacts and potential risks of the proposed system
14
15 Customer satisfaction  Analyzed existing problems at the airport and identified the business requirements to ensure
ma
16 the proposed system meets the airport’s expectations
17  Assessed and disclosed any potential risks found
18  Continuously provided any findings to the airport during the development process
19
20 Operational compliance  Worked with the airport’s IT and surveillance teams to understand the airport’s regulations
tio

21 and policies when accessing the cameras and dealing with the data
22  Prior to data collection, obtained human ethics clearance from both parties to ensure the
23 proposed AI solution is compliant with existing regulations and policies and operates lawfully
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
Page 33 of 44 Information Technology & People

Table VIII. List of actions taken to establish and maintain trust in the Develop and Deploy the solution stages
1
2 Checklist for Develop and Actions/Strategies taken
3 Deployment
4
5 Reliable operations  Conducted all experiments and testing in the internal airport’s machine
6  Evaluated the AI models’ performance through rigorous benchmarking with state-of-the-art-
7 model, compared it with existing systems and manual observation prior to deployment
8  Deployed the system within a secure network hosted by the airport and only performed
9 computations in the airport’s network. No data transmission outside the airport’s network
Inf
10
11 Empowered employees  Obtained continuous feedback from the airport’s personnel to improve the proposed system’s
12 usefulness
or
13
14 Engaged customers  Tailored the AI solution to the airport’s context and requirements. Further fine-tuning
15 model and continuous training were conducted to meet the expected performance

ma
16 Provided continuous evidence of the benefits of the AI solution. This can be seen from the
17 evolution of AI adoption from crowd counting to social distancing breach detection
18
19 Transformed products  Demonstrated the proposed AI solution to the airport management to introduce a new
20 approach to ensuring passengers’ safety and comfort at the airport through surveillance while
tio

21 respecting people’s privacy

22
23
nT

24
25
26
27
ec

28
29
30
hn

31
32
33
34
olo

35
36
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60
 Information Technology & People Page 34 of 44

Table IX. List of actions taken to establish and maintain trust in the Operate, Monitor, and Scale-up stages
1
2 Checklist for Operate, Actions/Strategies taken
3 Monitor and Scale-up
4
5 Continuous refinement  Enforced continuous development of AI through rigorous evaluation and benchmarking
6 with the-state-of-art AI model
7  Evaluated the AI performance with existing systems to ensure accuracy in the decision-
8 making process
9  Obtained continuous feedback from the airport’s personnel to improve the system’s
Inf
10 accuracy
11  Published the work in internationally peer-reviewed journals and presented the demo at
12 international events to justify and evaluate the solution, and obtain experts’ feedback
or
13
14 Responsible business practice  Followed the airport’s regulation and policies during data collection and experiments
15  Airport applied strict monitoring to ensure any activities during system development did
ma
16 not breach the agreement
17
18 Individual/user trust  Applied AI solution aims to support the airport’s personnel in making well-informed
19 decisions. The AI solution does not make automatic decisions and replace existing
20
tio

systems
21  Introduced the AI solution to reduce human involvement. Human is still required to make
22 final decisions (automatic notifications will be sent to personnel when prompt responses
23 are needed)
nT

24  Conducted periodic evaluation through in-person meetings to discuss existing issues

25
 Documented any agreements between both parties and presented regular updates and
26
results to promote transparency
27
 Applied strict data retention policy by keeping video feeds for up to 7 days
ec

28
29  Only collected data related to the objectives
30
Societal (community) trust  Conducted the research in full accordance with human research ethics committee and
hn

31
legal department’s requirements
32
33  Ensured that all camera access is approved by the airport’s surveillance department prior
34 to access. Any changes were reported and documented
olo

35  Informed the public by placing signage stating that video observation was in place to
36 maintain passenger awareness
37
38
gy

39
40
41
42
&P

43
44
45
46
47
eo

48
49
50
ple

51
52
53
54
55
56
57
58
59
60