SOCIAL DISTANCE
DETECTOR USING
YOLO V4
LUTFIYAH NUR FADHILAH
3120640018
D4 LJ T.Informatika
Tahun Ajaran 2020/2021
ADVISOR
Tita Karlita S.Kom., M.Kom. 1979101420021220022
M. Udin Harun Al Rasyid,Ph.D 198108082005011001
BACKGROUND.
Coronavirus disease 2019 or COVID-19 has infected hundreds of
millions of people and claimed thousands of millions of lives worldwide.
The spread of the COVID-19 virus can also be said to be quite fast
so that it has an impact on the order of life in society
The government has taken various ways to reduce the number of
COVID-19 cases, especially in Indonesia, one of which is to
implement social distancing while on the move
STATEMENT OF THE PROBLEM.
During the COVID-19 pandemic, the government has tried
to implement various social distancing practices, such as
limiting travel between cities, controlling regional borders,
closing schools and entertainment venues. However,
monitoring the number of infections spread and the
efficiency of barriers is difficult. People have to go out and
meet essential needs such as necessary food, health care
and jobs.
RESEARCH OBJECTIVE.
Applying the concept of Computer Vision to create a monitoring
system or surveillance and their capabilities for automatic
detection and recognition of selected parameters of people using
the YouOnlyLookOnce v4 . Human objects will be monitored
whether they meet the safe distance according to WHO
requirements. When the tool detects a decrease in the percentage
of implementing Social Distancing, then system will result the
output (Red Boundingbox) and the number of people at risk. The
author claims this method can support reducing the spread of
COVID-19.
Created an artificial intelligence system that
focuses on detecting human objects and
calculating the distance between objects in a
crowd by using the calculation of the
DELIMITATION Eucladian distance between the centroid of
OF the bounding box.
THE
The method used in this final project is the
PRODUCT YouLookOnlyOnce (YOLO) method which
uses the Google Open Image dataset. The
dataset is used to fine-tune YOLOv4 and
Pretrained YOLOv3 models as initial weights.
RESEARCH
METHOD
DESIGN SYSTEM
DATA
CHARACTERISTIC
Open image is a dataset of ~9 million
images annotated with image-level labels,
object bounding boxes, object
Dataset
segmentation masks, visual relationships.
In this study, the author uses the weight of the pre- Contains a total of 16 million bounding
trained Yolov4 model that has been trained with the boxes for 600 object classes on 1.9 million
MSCOCO dataset. The weight of the pre-trained
images, making it the largest dataset
model was used as the initial weight of the training.
The process of training the model with the initial available with local object annotations. The
weight values derived from the pre-trained model is boxes have been mostly manually drawn by
called Fine Tuning. The author performs finetuning professional annotators to ensure accuracy
on the pretrained model using the Google and consistency.
OpenImage dataset.
cr:https://storage.googleapis.com/openi
mages
FINE TUNING
A lot of effort in solving any machine learning problem goes into preparing the
data.the author does a finetuning with the goal detection process only focusing
on the objects needed. this can speed up every process for object detection
which will be continued with the distance detection.
The author uses Jupyter Notebook (ML AZURE) to Training OpenImage
Dataset . This process uses one class only (Human_Body) , 192 epochs, 1000
training data, 1000 testing data and learning rate 0.001. The training time
required for this training model is 26 hours. the mAP result is 44.7% . with this
value the detection process is three times faster, this can be seen from the
FPS.
Source: https://storage.googleapis.com/openimages
THE
RESULT
The Result
displays the detection results of the CCTV camera
capture. In the upper corner, there is information
about the number of people who are at risk of
spreading the virus due to the violation of the safe
distance limit. In accordance with the system
algorithm, the violating object will have a red
bounding box and when it is at a safe distance it will
have a green bounding box. . It aims to compare
the area to get information on the distance
detection area. The video also produces a display
of the distance between objects when a violation is
detected. Each detected object centroid will draw a
line and display the distance number in centimeters
(cm).
RESULT OF OBJECT DETECTION
the true positive parameter
indicates the system can detect
human objects in the image with the
right location of the bounding box.
Of the total images, as many as 91
images that were detected were
correct so that they got an accuracy
value 0.938.
RESULT OF DISTANCE MEASURMENT
The difference in the results on the error
value is influenced by the constant
calculation by calibrating the camera to
calculate the distance. From the results of
the error calculation, it can be seen that the
error value of reading the image processing
results and the actual results are still in the
small error range, namely the largest error
value is at 21.6%. This value is a
comparison value which is very far from
100%. So it can be said that the distance
detection results in this study have a
relatively high level of truth.
DISCUSSION.
1. The results of the training process using the pre-trained model
YOLOV4 get the best model with a batch size of 64 and a learning rate
of 0.001 spending ± 26 hours of learning time and getting an average
mAP of 44.7% and a total loss value of 2.452.
2. The results of human object detection using Yolov4 can detect
human objects with a bird's eye view. In addition, the system is able to
detect human objects with object images even though the objects
caught on the screen pass the camera frame limits.
OILOFTROP GNILLEDOM
3. The results of human object detection with moving object conditions
get an accuracy of 0.938. This value is good enough for human
detection which is then continued for distance detection.
4. The system has an average error in the calculation of the distance in
the image with the actual distance, which is 21.6%. This value is not
good enough to calculate the safe distance for human objects in the
effort of social distancing protocols.
MONITORING COVID-19 SOCIAL DISTANCING WITH PERSON DETECTION AND
TRACKING VIA FINE-TUNED YOLO V3 AND DEEPSORT TECHNIQUE
Narinder Singh Punn, Sanjay Kumar Sonbhadra and Sonali (May 2020)
DEEPSOCIAL: SOCIAL DISTANCING MONITORING AND INFECTION RISK
ASSESSMENT IN COVID-19 PANDEMIC.
Mahdi Rezael and MohsenAzarmi (October 2020)
A DEEP LEARNING-BASED SOCIAL DISTANCE MONITORING FRAMEWORK FOR
References COVID-19
Imran Ahmed and Misbah Ahmad (November 2021)
A VISION-BASED SOCIAL DISTANCING AND CRITICAL DENSITY DETECTION
SYSTEM FOR COVID-19
Dongfang Yang, Ekim Yurtsever, Vishnu Renganathan, Keith A.
Redmill, and Umit Ozguner (July 2020)
SOCIAL DISTANCING DETECTION SYSTEM WITH ARTIFICIAL INTELLIGENCE
USING COMPUTER VISION AND DEEP LEARNING
Vinitha. V and Velantina (Ausgust 2020)
THANK YOU
All Source : Writer
