Machine Learning-Based Vehicle Intention Trajectory

Recognition and Prediction for Autonomous Driving

Hanyi Yu1* Shuning Huo2
Computer Science Statistics
University of Southern California Virginia Tech
Los Angeles, CA, USA Blacksburg, VA, USA
*
Corresponding author: e-mail:hanyiyu@usc.edu shuni93@vt.edu

Mengran Zhu3
Computer Engineering Yulu Gong4

Miami University Computer & Information Technology

Oxford, OH USA Northern Arizona University
mengran.zhu0504@gmail.com Flagstaff, AZ, USA
yg486@nau.edu

Yafei Xiang5
Computer Science
Northeastern University
Boston, MA, USA
xiang.yaf@northeastern.edu

Abstract—In recent years, the expansion of internet has ignored the human factor, which is very important for the
technology and advancements in automation have brought accurate modeling of autonomous lane change decision. In this
significant attention to autonomous driving technology. Major paper, a new autonomous lane change decision model is
automobile manufacturers, including Volvo, Mercedes-Benz, and proposed by integrating the human factors represented by
Tesla, have progressively introduced products ranging from driving style. The model takes into account not only the
assisted-driving vehicles to semi-autonomous vehicles. However, surrounding traffic information, but also the driving style of the
this period has also witnessed several traffic safety incidents surrounding vehicles to make lane change/lane keeping
involving self-driving vehicles. For instance, in March 2016, a
decisions. In addition, the model can mimic the decision-
Google self-driving car was involved in a minor collision with a
making operations of a human driver by learning the driving
bus. At the time of the accident, the autonomous vehicle was
attempting to merge into the right lane but failed to dynamically
style of a self-driving car. The results show that the model can
respond to the real-time environmental information during the accurately describe the human decision-making strategy,
lane change. It incorrectly assumed that the approaching bus simulate the human driver's lane change action, and the
would slow down to avoid it, leading to a low-speed collision with prediction accuracy can reach 98.66%.
the bus. This incident highlights the current technological In traditional mathematical lane change decision-making
shortcomings and safety concerns associated with autonomous models, some of them focus on describing the overall impact of
lane-changing behavior, despite the rapid advancements in
lane change on traffic flow, ignoring the consideration of
autonomous driving technology. Lane-changing is among the
human drivers. Except for the lane change decision model
most common and hazardous behaviors in highway driving,
significantly impacting traffic safety and flow. Therefore, lane- based on game theory, most models treat the lane change
changing is crucial for traffic safety, and accurately predicting decision process as a single driver decision, and do not capture
drivers' lane change intentions can markedly enhance driving the interaction between lane change vehicles and surrounding
safety. This paper introduces a deep learning-based prediction vehicles. The rules and parameters defined in most existing
method for autonomous driving lane change behavior, aiming to models are limited and few in number, which cannot accurately
facilitate safe lane changes and thereby improve road safety. describe lane change behavior. The lane change decision time
is often incorrectly labeled as crossing time. The above
Keywords-Autonomous driving; CNN-LSTM model; Deep problems lead to the poor prediction accuracy of the traditional
learning; Autonomous lane change lane change decision model. To overcome the limitations of
traditional lane change decision models, an autonomous lane
I. INTRODUCTION change decision model (DSA-DLC) that considers driving style
The purpose of autonomous lane change is to achieve faster perception was proposed. The model implicitly represents the
driving speed or better driving conditions. Most existing driving style of vehicles by extracting it from the driving
research on the modeling of autonomous lane change decisions operation graph (DOP) of historical trajectories, and the driving
style of surrounding vehicles and autonomous vehicles was
taken as human factors. The relationship between lane change networks (CNNS) to a series of visual images, the motion of a
decision, traffic factor and human factor is modeled by neural simple physical system is predicted directly from image pixels.
network, and the autonomous lane change decision model is Later, researchers proposed a system in which CNNs were used
constructed to predict the autonomous lane change decision of to predict short-term vehicle trajectories, with BEV raster
human driver. images encoding the surroundings of a single participant as
input, which were then also applied to vulnerable traffic
To solve the above problems, this paper proposes a CNN- participants.Despite the success of these methods, they do not
LSTM lane change intention prediction model based on deep address the underlying multimodal problem of possible future
learning, which combines the advantages of convolutional trajectories required for accurate long-term traffic prediction.
neural network (CNN) and long short-term memory network
(LSTM) to improve road traffic safety and fluency. The model At present, there are many researches to solve the problem
first processes high-dimensional spatiotemporal traffic data of multimodal modeling. Mixed density networks (MDNs) are
through CNN to extract key features, including vehicle speed, traditional neural networks that solve multimodal regression
acceleration, relative position and other information. problems by learning the parameters of Gaussian mixture
Subsequently, the LSTM uses these features to grasp the time- models. However, MDNs is often difficult to train in practice
series nature of traffic flow and capture the long-term due to numerical instability when operating in high-
dependencies in driving behavior. This combination enables dimensional Spaces. Therefore, based on the above practical
the model to effectively predict the driver's lane change problems, the method proposed in this paper directly calculates
intention and provide more accurate decision support for the prediction results of multiple modes on a single forward
intelligent transportation systems and autonomous driving CNN model.
technologies. On this basis, future studies can further explore
the adaptability of the model in different traffic scenarios and C. Construction of lane change decision model
how to optimize the model performance through real-time data
feedback.
II. RELEVANT THEORIES AND MODELS

A. Behavioral prediction in autonomous driving systems

The behavior prediction in the automatic driving system is
to accurately detect the current position, speed, direction of
movement and other information of the surrounding vehicles
through various sensors installed by the autonomous vehicle
itself, and then use it to predict its future trajectory. In recent
years, vehicle trajectory prediction methods are mainly divided Figure 1. DSA-DLC decision model architecture
into two categories: reinforcement learning model based
method and neural network-based interactive learning and y is the lane change decision vector, which includes three
social perception method. Model-based methods use elements corresponding to the probabilities of lane keeping,
appropriate learning algorithms to achieve results in specific lane changing left and lane changing right respectively.
scenarios. In recent years, the method based on the game model Considering the lane change incentive factor, safety factor,
has often been used for trajectory prediction in intelligent tolerance factor and driving style extracted from DOP D, the
vehicles. Li et al. The CMMetric algorithm used for behavior decision-making formula is as follows:
modeling and prediction is used to sort the CMetric values of
the surrounding vehicles to determine the passing order of the y = fLC Fincentive ，FSAFETY ，Ftolerance ，FDS D (1)

�� ，�� ，� ∈ �, �, � , � , 퐹 , 퐹 , 퐴� , 퐴�
surrounding vehicles and predict the trajectory. The above two Hypothesis 1:
methods based on the game model are aimed at the game
between the two players, and can not meet the trajectory
prediction of the target vehicle in many scenarios. (2)
B. Machine learning predictive models Where v is the speed of vehicle i and d is the longitudinal
The success of deep learning in many practical applications distance from vehicle i to vehicle I.
has prompted research into its application in motion prediction.  Lane change facilitation factor

퐹��������� = 퐹� �� − �� ， �� − �� ， �� −
With the recent success of recurrent neural networks (RNNS),
one line of research called short-time memory (LSTM) has
�� ， �� − �� ， �� − ��
been used for sequence prediction tasks. The authors used
LSTM to predict the future trajectory of pedestrians in social
interactions. In the literature, LSTM has been applied to predict (3)
vehicle position using past trajectory data. Another RNN
variant, called a gated recursive unit (GRU), is combined with  (2) Lane change safety factor
conditional variational autoencoders (CVAE) to predict vehicle
trajectory. In addition, by applying convolutional neural
 퐹������ = �� �퐹 ，�퐹 ， �� − �퐹 ， �� − of the pedestrian body, such as the shoulder, and applying the

�퐹
Gauss process dynamics model based on balance. Based on the
(4) social force model, Rinke et al. proposed a multi-level
description method of road user movement and its interaction,
The risk of collision is affected by the distance and speed and discussed pedestrian movement target points and possible
difference between the vehicle and the vehicle behind the target trajectories in layers. By first determining the target point,
lane. When the distance between the self vehicle and the rear using Lagrange polynomial to estimate other trajectories in turn,
vehicle is large enough, and the speed of the self vehicle is and then using the conflict avoidance strategy based on social
higher than the speed of the rear vehicle in the target lane, the force model to select trajectories, the best predicted trajectories
collision risk is low. are generated.
 (3) Tolerance factor
퐹��������� = �� �� − �� ·�ℎ
(5)
It is used to measure the driving condition in the current
lane and is highly correlated with the safe headway of the car
and the distance of the car in front. Depending on driving
experience, drivers may be more inclined to stay farther than a
safe distance on the highway to improve driving comfort.
Among them, f_I, f_s and f_T are constructed by neural Figure 2. Framework of trajectory prediction network for trajectory ++
network as traffic factors input to the lane change decision
model. However, model-based methods are highly dependent on
theories and scenarios, and require a large amount of data to
 (4) Model design train models. Therefore, in recent years, more researchers are
still committed to methods based on the combination of neural
y = fLC Fincentive ，FSAFETY ，Ftolerance，FDS D =
�� − �� ，�� − �� ，�� − �� ，
networks and attention mechanisms, generally
RNN/LSTM/GRU and other recurrent neural networks and
fθ �� − �� ，�� − �� ，�퐹 ，�퐹 ，�� − �퐹 ，
their variants to combine attention mechanisms. Zhou et al. use

�� − �퐹 ，�� − �� ·�ℎ ，�
recurrent neural Networks (RNN) and Graph Convolutional
Networks (GCN) to simulate the state of pedestrians and their
interactions by considering the movement information of
(6)
individual pedestrians and their interactions with surrounding
theta is a set of parameters of the model, and the lane pedestrians (Figure 3). The trajectory is predicted by
change decision is a multi-parameter nonlinear problem, which combining attention mechanism and companion loss function.
is solved by deep learning. Using trajectory data set, input
variables and decisions are extracted frame-by-frame and
modeled as supervised learning problems. The learning goal is
to find a model to minimize the long-term average loss,
calculate the loss value and optimize the neural network,
defined as:
−1
yi, fθ .
1 N N 3
N i=1 L = N i=1 c=1 yic log yic (7)
Figure 3. SGCN network framework

The DSA-DLC decision model consists of convolutional Liu et al. proposed a prediction model of whether
neural network (CNN) and fully connected neural network pedestrians cross the road based on graph convolution, and
(FC). CNN has shown a strong image classification ability in predicted pedestrian trajectory within a future time range by
capturing the hidden features of images. The driving style is reasoning only about the relationship between pedestrians and
regarded as the hidden characteristics of the vehicle, and CNNs the surrounding environment and their own body movements.
is used to extract the hidden driving style from the DOP. Since Shi et al. proposed a novel analytic Graph Convolution
the driving styles of self-driving cars and surrounding cars have Network (SGCN), which combined Sparse directed interaction
different effects on lane change decisions, two CNNs are used, with motion trends for pedestrian trajectory prediction. The
one for the self-driving car DOP and the other for the model used a sparse graph learning method, based on an
surrounding car DOP. attention mechanism, to score track points. It is fed back into
D. Neural networks and attention mechanisms the asymmetric convolutional network to obtain high-level
interaction features. The adjacency matrix obtained after
Model-based methods include Gauss process model and normalization can represent the sparse graph. Finally, the
social force model. Minguez et al. proposed a method to predict trajectory is predicted by combining the parameters of double
pedestrian trajectory by collecting the information of key parts Gaussian distribution estimated by the graph convolutional
network for trajectory prediction. Compared with previous that time period. We mark the start and end of this lane change
predictive modeling methods, it makes targeted selection of trajectory when θ reaches the boundary value θbound :|θ|=
interactive pedestrian information, vehicle information and θbound.
environmental information, instead of directly applying all the
above information within a certain range to modeling without
difference.
In conclusion, The main advantage of the method based on
deep learning neural network and long short-term memory
network (LSTM) to predict lane change behavior of
autonomous driving is that it can effectively process and
analyze time series data, so as to accurately predict lane change
behavior of vehicles in different traffic situations. By learning Figure 4. The start, change point and end point of the lane change trajectory
from historical data, this method predicts future driving paths
and potential risks in real time, supporting autonomous driving Figure 5.(a) The lane change prediction point is determined
systems. It plays a crucial role in enhancing the safety and if the vehicle is predicted to enter the lane change for 3
efficiency of autonomous vehicles. By predicting the potential consecutive time steps. The lane change prediction time is
behavior of surrounding vehicles, the autonomous driving defined as the time interval between the lane change point and
system can make adjustments in advance or take avoidance the lane change prediction point. (b) n continuous time steps
measures to avoid traffic accidents and ensure the safety of are packaged into a trajectory segment. If the NTH time step of
passengers and pedestrians. In addition, this method helps to the track segment is the lane following time step, then the
improve the smoothness of traffic flow, reduce unnecessary segment is a lane following segment, otherwise it is marked as
lane changes and braking, and thus improve road utilization a lane changing segment.
and driving efficiency.
III. METHODOLOGY
A. Data extraction and processing
 Data set
The data set in this paper is primarily derived from the
Federal Highway Administration's Next Generation Simulation
(NGSIM) dataset for extracting vehicle trajectories and
modeling lane change predictions, which has been adopted by
many previous studies. In a time interval of 0:1 second, the
dataset recorded the position, speed, acceleration, and headway
information of each vehicle on U.S. highways 101 and
Interstate 80 (I-80). Both locations contain 45 minutes of
vehicle trajectory data. Highway 101 is 640 meters long and
has five main lanes and six service lanes, while I-80 is about
500 meters long and has six main lanes. Figure 5. Vehicle behavior marks track segments

In this paper, six vehicle trajectory data sequences were Figure 5 (b) depicts the way we label track segments. For
extracted from NGSIM, each of which lasted 10 minutes. We each vehicle, n consecutive time steps are packaged into a track
removed the first 5 minutes from each 15-minute sequence to segment. If the NTH time step of the track segment is the lane
ensure a sufficient number of vehicles in each frame. For each change time step, then the segment is a lane change segment,
sequence, the first 2 minutes are defined as the test set and the otherwise it is marked as a lane following segment. In this
remaining 8 minutes are defined as the training set. Since the article, we set n to 6,9, and 12 to determine the effect of the
data is recorded at 10 frames per second, we can get a total of length of the historical track on the final result.
1,200 test time steps and 4,800 training time steps.
In summary, this paper can obtain about 60,000 lane
 Data tag changes, plus 400,000 cars for training. This obviously
involves a problem of data imbalance, in which there are many
Vehicles are marked as "intending to change lanes to the more lane following pieces than lane changing pieces used for
left," "intending to travel along the lane," or "intending to training, which will lead to overfitting during training. To solve
change lanes at each time step." The way we mark the status of this problem, we randomly selected the same number of
the vehicle is as follows. fragments N from the change-left pool, the change-after pool,
As shown in Figure 4, we first collect all lane exchange and the change-right pool and mixed them together as the
points, i.e. the point at which the vehicle gravity point crosses training data set. To maximize the use of data, N is set to the
the dotted line dividing the lane, the vehicle. If the vehicle is at number of pieces (30,000 pieces) in the pool on the right side
the lane change point at time step t, we check its trajectory at of the lane change.
[t-δt, t+δt] (δt=2s) and calculate its heading value θ for
 Then, given the first (n1) time step historical track and
neighbor information in the test set, the lane change intent of
each vehicle is predicted at each time step. The predicted time
of lane change is also calculated after filtering the results.
Specifically, the lane change prediction point is determined if
the forecast vehicle makes 3 continuous time steps of lane
change, and the lane change prediction time is defined as the
time interval between the lane change point and the lane - the
change prediction point, as shown in Figure 5 (a).
B. Method
In this paper you need to predict whether the car will
Figure 6. LSTM network structure for lane change intent prediction
change lanes and which lane it will merge into. We use an
LSTM to enable the agent to reason about the vehicle's As shown in Figure 6, we adopt the LSTM network
historical trajectory information. However, since human architecture to deal with this problem of intention prediction of
decision-making behavior will also depend on the surrounding lane change. The embedded dimension selected for the features
vehicles, we also use vehicle neighborhood information as of the vehicle itself and its neighbors is 64, and the hidden
input to the network. dimension of the LSTM network is 128. We chose a learning
 Input function rate of 0:00 0125 and used the soft-max cross-entropy loss as
the training loss: loss =-Σi=1yi´ log (yi). Where y is the true
Two types of input features are used for the prediction
label of the intention of the i lane change (yi ´ =1, intention
algorithm:
exists, yi´=0, intention does not exist. i∈{1,2,3}. y1´ is the left
(1) Information about the vehicle itself and (2) information intent to change lanes, y2´ is the intent to follow lanes, and y3´
about the vehicle's neighbors. Information about the vehicle is the right intent to change lanes). yi is the predicted output
itself includes: probability of the model with the i lane change intention after
a) Vehicle acceleration passing through the soft-max layer.
b) The steering Angle of the vehicle relative to the road
c) Global lateral vehicle position relative to the lane
d) Global longitudinal vehicle position relative to the
lane TABLE I. COMPARISON OF THE ACCURACY OF CHANGING LANE
PREDICTIONS
Vehicles estimating their lane change intentions are
provided by the following features: Real Predict Left Following Right
SA-LSTM Left 87.40% 12.34% 0.26%
e) Presence of the left lane (1 if present, 0 otherwise) Following 7.47% 85.33% 7.20%
f) Presence of the right lane (1 if present, 0 otherwise) Right 2.94% 11.22% 85.84%
Feedforw
g) Longitudinal distance between self vehicle and left ard
Left 84.6% 15.40% 0%
Following 2.61% 83.78% 13.61%
front vehicle Neural Right 2.44% 12.91% 79.65%
Network
h) The longitudinal distance between the self vehicle and
Logistic Left 64.91% 35.03% 0.06%
the vehicle in front Regressio Following 9.88% 82.87% 7.25%
i) Longitudinal distance between self vehicle and right n Right 0.05% 36.30% 63.65%
front vehicle C. Test result
j) Longitudinal distance between self vehicle and left Comparisons with other network structures are made with
rear vehicle feedforward neural networks, logistic regression, and LSTMS
k) Longitudinal distance between self vehicle and rear without adjacent feature inputs to show the advantages of
vehicle adding historical tracks and environmental factors. Table I and
l) Longitudinal distance between self vehicle and right Figure 6 show the classification accuracy calculated by our
rear vehicle algorithm, feedforward neural network and logistic regression.
 Network structure The method we call Environment Aware (SA) -LSTM, based
on the advantages of historical track information and neighbor
information, outperforms the other two methods in terms of
prediction accuracy on all classification types (left transition
track, right transition track, and right transition track).
D. Comparison of different trajectory lengths
Set the historical track lengths of the LSTM structure to 6,9,
and 12, and compare them to each other. The results are shown
in Table II and visualized in Figure 7. We compared the results
of five different trajectory sequences to help us get a general information, the model provides a comprehensive
idea of the trend of curve change. In all prediction scenarios, understanding of lane change behavior, significantly
the prediction accuracy increases as the history length increases outperforming traditional approaches in prediction accuracy.
(left lane change, right lane change).
B. Future Prospects of Deep Learning in Autonomous
Driving and Beyond
The success of the CNN-LSTM model in predicting lane
change intentions opens up new avenues for the application of
deep learning in autonomous driving and other fields. The
ability to process and analyze large datasets to predict complex
behaviors has far-reaching implications, from improving road
safety to optimizing traffic flow and reducing congestion.
Future research can explore the adaptability of this model in
different traffic scenarios and its integration into real-time
autonomous driving systems. Beyond autonomous driving, the
principles and methodologies developed in this study can be
applied to other domains where predicting human behavior is
crucial, such as robotics, smart cities, and personalized services.
Figure 7. Comparison of prediction accuracy of different methods The convergence of deep learning, big data, and computational
power promises to revolutionize how we understand and
Based on the accuracy of the different methods, A-LSTM interact with the world around us, driving innovation and
outperformed the other two types in all classification types, improving quality of life across various sectors.
including right lane change, lane follow, and left lane change.
Therefore, it can be seen that the longer the historical track REFERENCES
length, the more information we can obtain from the previous
track, and the higher the accuracy in the final result. However, [1] D. Clarke, J. W. Patrick, and J. Jean,“Overtaking road-
there is a trade-off between length and computation time. As accidents:Differences in manoeuvre as a function of driverage,”
Accident Analysis & Prevention, vol. 30, no. 4, pp. 455–467, 1998.
the length increases, the improvement in accuracy slows down.
From the figure, we can see that length = 12 achieves only [2] Q. Yang and N. K. Haris, “A microscopictraffic simulator for evaluation
slightly higher accuracy compared to length = 9. More of dynamic traffic management systems,”Transportation Research Part C:
importantly, the length of the historical trajectory should not be Emerging Technologies, vol. 4, no. 3, pp.113–129,1996.
too long. [3] C. R. Baker and J. M. Dolan, “Trafficinteraction in the urban challenge:
Putting boss on its best behavior,” inIEEE/RSJ International Conference
IV. CONCLUSION on Intelligent Robots and Systems. IEEE,2008.
[4] Liu, T., Xu, C., Qiao, Y., Jiang, C., & Yu, J. (2024). Particle Filter
SLAM for Vehicle Localization. arXiv preprint arXiv:2402.07429.
The paper presented an innovative approach to enhancing [5] J. Liao, A. Kot, T. Guha and V. Sanchez, "Attention Selective Network
the safety and efficiency of autonomous driving through a deep For Face Synthesis And Pose-Invariant Face Recognition," 2020 IEEE
learning-based prediction model for lane-changing behavior. It International Conference on Image Processing (ICIP), Abu Dhabi,
United Arab Emirates, 2020, pp. 748-752, doi:
underscores the importance of incorporating human factors, 10.1109/ICIP40778.2020.9190677.
represented by driving styles, into autonomous lane change [6] J. Liao, V. Sanchez and T. Guha, "Self-Supervised Frontalization and
decision models. Traditional models often overlook the Rotation Gan with Random Swap for Pose-Invariant Face Recognition,"
interaction between vehicles and the dynamic nature of driving 2022 IEEE International Conference on Image Processing (ICIP),
behaviors, leading to inaccuracies in predicting lane change Bordeaux, France, 2022, pp. 911-915, doi:
decisions. By integrating CNN and LSTM networks, the model 10.1109/ICIP46576.2022.9897944.
successfully captures the complex dynamics of traffic flow and [7] Z. Shangguan, Y. Zhao, W. Fan and Z. Cao, "Dog Image Generation
using Deep Convolutional Generative Adversarial Networks," 2020 5th
driver behavior, offering a more accurate and reliable International Conference on Universal Village (UV), Boston, MA, USA,
prediction mechanism. This approach not only addresses the 2020, pp. 1-6, doi: 10.1109/UV50937.2020.9426213.
limitations of previous models but also sets a new standard for [8] J. Wei, J. M. Dolan, and B. Litkouhi, “ A prediction and cost function-
research in autonomous driving technologies. based algorithm for robust autonomous freewaydriving. ” in Intelligent
Vehicles Symposium (IV). IEEE, 2010.
A. Experimental Conclusion Analysis [9] Yu, Liqiang, et al. “Research on Machine Learning With Algorithms and
The experimental results demonstrate the model's superior Development”. Journal of Theory and Practice of Engineering Science,
performance in predicting lane change intentions with a vol. 3, no. 12, Dec. 2023, pp. 7-14, doi:10.53469/jtpes.2023.03(12).02.
remarkable accuracy rate of 98.66%. The use of CNN-LSTM [10] Xin, Q., He, Y., Pan, Y., Wang, Y., & Du, S. (2023). The
implementation of an AI-driven advertising push system based on a NLP
architecture allows for the effective processing of algorithm. International Journal of Computer Science and Information
spatiotemporal traffic data, capturing the nuanced patterns of Technology, 1(1), 30-37.0
driving behavior over time. The comparison with other network [11] Zhou, H., Lou, Y., Xiong, J., Wang, Y., & Liu, Y. (2023). Improvement
structures, such as feedforward neural networks and logistic of Deep Learning Model for Gastrointestinal Tract Segmentation
regression, further validates the model's effectiveness. By Surgery. Frontiers in Computing and Intelligent Systems, 6(1), 103-
incorporating the driving style and surrounding traffic 106.6
[12] Implementation of an AI-based MRD Evaluation and Prediction Model [18] Xiaofeng Li,Shijing Wu,Xiaoyong Li,Hao Yuan,Deng Zhao.Particle
for Multiple Myeloma. (2024). Frontiers in Computing and Intelligent Swarm Optimization-Support Vector Machine Model for Machinery
Systems, 6(3), 127-131. https://doi.org/10.54097/zJ4MnbWW. Fault Diagnoses in High-Voltage Circuit Breakers[J].Chinese Journal of
[13] Zhang, Q., Cai, G., Cai, M., Qian, J., & Song, T. (2023). Deep Learning Mechanical Engineering,2020,33(1):104-113.
Model Aids Breast Cancer Detection. Frontiers in Computing and [19] Li Mei, NING Dejun, Guo Jiacheng. CNN-LSTM model based on
Intelligent Systems, 6(1), 99-102.3 Attention mechanism and its Application [J]. Computer Engineering and
[14] Manchini Carlos, et al. "A new approach to data differential privacy Applications,2019,55(13):20-27. (in Chinese)
based on regression models under heteroscedasticity with applications to [20] D. Kasper, G. Weidl, T. Dang, G. Breuel,A. Tamke, A. Wedel, and W.
machine learning repository data." Information Sciences 627. (2023): Rosenstiel, “Object-oriented bayesian networks fordetection of lane
[15] Xu, J., Pan, L., Zeng, Q., Sun, W., & Wan, W. (2023). Based on change maneuvers,” IEEE Intelligent Transportation Systems
TPUGRAPHS Predicting Model Runtimes Using Graph Neural Magazine,vol. 4, no. 3, pp. 19–31, 2012.
Networks. Frontiers in Computing and Intelligent Systems, 6(1), 66- [21] G. Weidl, A. L. Madsen, V. Tereshchenko,W. Zhang, S. Wang, and D.
69.7 Kasper, “Situation awareness and early recognition oftraffic maneuvers,”
[16] Wan,Weixiang, et al. "Development and Evaluation of Intelligent in 9th EUROSIM Congress on Modelling and Simulation. IEEE,2016.
Medical Decision Support Systems." Academic Journal of Science and [22] C. Dong, J. M. Dolan, and B. Litkouhi, “ Intention estimation for ramp
Technology 8.2 (2023): 22-25. merging control in autonomous driving,” inIntelligent Vehicles Symposium
[17] Tian, M., Shen, Z., Wu, X., Wei, K., & Liu, Y. (2023). The Application (IV), 2017 IEEE. IEEE, 2017, pp. 1584–1589.
of Artificial Intelligence in Medical Diagnostics: A New
Frontier. Academic Journal of Science and Technology, 8(2), 57-61.7