Surrounding vehicle track prediction method based on deep learning algorithm
The surrounding vehicle trajectory prediction network constructed through deep learning algorithms solves the problem of insufficient vehicle trajectory prediction accuracy in complex traffic scenarios, realizes accurate prediction of surrounding vehicle trajectory, and improves the safety of autonomous driving and the operation efficiency of the traffic system.
Patent Information
- Application Number
- CN202510679532.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-26
- Publication Date
- 2025-08-08
AI Technical Summary
The existing vehicle trajectory prediction methods are insufficient in complex traffic scenarios, especially in multi-lane expressways and other scenarios, making it difficult to accurately predict the vehicle's lane change behavior, resulting in insufficient prediction accuracy and reliability.
Using a peripheral vehicle trajectory prediction method based on deep learning algorithms, data is collected using lidar and millimeter microwave radar, and a trajectory prediction network including Bi-LSTM, Softmax layer, LSTM, channel attention layer, time attention layer and multi-scale feature fusion module is optimized, and the trajectory prediction model is optimized by combining the sliding window method and the Adam optimizer for training.
It improves the accuracy and generalization ability of trajectory prediction, can accurately predict the trajectory of surrounding vehicles, ensure that autonomous vehicles drive safely and efficiently in complex traffic environments, reduce the risk of traffic accidents, and improve the operation efficiency of the traffic system.
Smart Images
Figure CN120452233A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent transportation systems, and in particular to a method for predicting surrounding vehicle trajectories based on a deep learning algorithm, which is applicable to road scenarios in the diversion area of a one-way three-lane expressway. Background Art
[0002] In modern intelligent transportation systems, vehicle trajectory prediction is a key technology for achieving autonomous driving, traffic flow optimization, and accident prevention. With the acceleration of urbanization and the continuous growth of traffic volume, predicting vehicle behavior in complex road environments has become increasingly important. However, existing vehicle trajectory prediction methods face numerous challenges when dealing with complex scenarios.
[0003] Traditional trajectory prediction methods primarily rely on physical models or simple statistical methods. While these methods can achieve some success in certain simple scenarios, their accuracy and reliability are often insufficient in complex traffic environments, such as multi-lane freeways and ramp entrances. For example, methods based on physical models assume that vehicle motion follows the laws of Newtonian mechanics. However, in real-world traffic scenarios, vehicle motion is influenced by numerous factors, such as driving habits, traffic signals, and road conditions, which are difficult to describe using simple physical models. Prediction models based on statistical methods, on the other hand, rely on extensive historical data and have limited generalization capabilities for new traffic scenarios or unusual conditions. In recent years, with the rapid development of deep learning technology, deep learning-based vehicle trajectory prediction methods have gradually gained attention. Deep learning algorithms can automatically learn complex patterns and regularities from large amounts of data, and possess stronger nonlinear fitting and generalization capabilities. However, existing deep learning-based trajectory prediction methods mostly focus on single lanes or simple scenarios. Research on complex scenarios such as multi-lane freeways is relatively limited. Furthermore, lane changes between vehicles are frequent and complex, making them even more difficult to predict. Summary of the Invention
[0004] This invention aims to address the problems of insufficient accuracy and weak generalization ability of existing vehicle trajectory prediction technology in complex traffic scenarios. It proposes a surrounding vehicle trajectory prediction method based on a deep learning algorithm. The purpose is to provide a reliable decision-making basis for autonomous driving vehicles through accurate trajectory prediction, thereby ensuring safe and efficient driving of vehicles in complex traffic environments, effectively reducing the risk of traffic accidents, and improving the overall operating efficiency of the transportation system.
[0005] In order to achieve the above-mentioned object, the present invention adopts the following technical solutions:
[0006] The present invention provides a method for predicting surrounding vehicle trajectories based on a deep learning algorithm, which is applied to a one-way expressway main line road scenario with at least two lanes. The method is characterized in that the method is performed as follows:
[0007] Step 1: Use LiDAR and millimeter-wave radar sensors to collect and preprocess the trajectory data of the main vehicle (ego) and surrounding vehicles to obtain the preprocessed trajectory data of the main vehicle (ego) and surrounding vehicles;
[0008] Step 2: Mark the pre-processed trajectory data of surrounding vehicles to obtain any Labels of surrounding vehicles at all times ;in, Indicates that the jth surrounding vehicle is Labels of moments;
[0009] Step 3: Use the sliding window method to perform window processing on the labeled surrounding vehicle trajectory data between time t1 and time t2 to obtain m feature subsequences ,in, represents the i-th feature subsequence;
[0010] Step 4: Build a trajectory prediction network for surrounding vehicles, including a behavior intention recognition module and a trajectory output module.
[0011] Step 4.1: Construct a behavioral intention recognition module, which includes a bidirectional long short-term memory neural network Bi-LSTM and a Softmax layer. Process and obtain The probability of the behavior intention of the surrounding vehicles ;
[0012] Step 4.2, build the trajectory output module, including: a long short-term neural network LSTM, a channel attention layer, a time attention layer, a multi-scale feature fusion module and a fully connected layer, and and Processing is performed to obtain trajectory prediction results ;
[0013] Step 5: Construct the total loss function Loss of the trajectory prediction network of surrounding vehicles;
[0014] Step 6: Use the Adam optimizer to train the trajectory prediction network of surrounding vehicles and continuously optimize the total loss function Loss to update the network parameters. When the total loss function Loss converges, the optimal surrounding vehicle trajectory prediction model is obtained, which is used to predict the trajectory of the input trajectory data.
[0015] The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to the present invention is also characterized in that, in step 1, the trajectory point of the pre-processed main vehicle ego at time t is ; Let the trajectory point of the jth surrounding vehicle at time t after the main vehicle ego preprocessing be recorded as , j=1,2,…,n, where T is the total sampling time, represents the lateral position of the jth surrounding vehicle at time t, represents the longitudinal position of the jth surrounding vehicle at time t, represents the speed of the jth surrounding vehicle at time t, represents the acceleration of the jth surrounding vehicle at time t; n represents the total number of surrounding vehicles, represents the lateral position of the vehicle ego at time t, represents the longitudinal position of the main vehicle ego at time t, represents the speed of the vehicle ego at time t, represents the acceleration of the vehicle ego at time t.
[0016] Furthermore, the step 2 includes:
[0017] Step 2.1: Calculate the heading angle of the jth surrounding vehicle at time t according to formula (1): ;
[0018] (1)
[0019] In formula (1), represents the lateral position of the jth surrounding vehicle at time t-1, represents the longitudinal position of the jth surrounding vehicle at time t-1;
[0020] Step 2.2: For the surrounding vehicles that have changed lanes, record the trajectory point that passes through the lane boundary at time t0 during the lane change process as the lane change point, and traverse along both sides of the lane change point:
[0021] When the heading angle value of the jth surrounding vehicle at time t1 satisfy When the trajectory point at time t1 is defined as the lane change starting point, the heading angle value at time t2 is satisfy , define the trajectory point at time t2 as the lane-changing end point;
[0022] For the surrounding vehicles that have not changed lanes, the trajectory point at time t1 is defined as the straight start point, and the trajectory point at time t2 is defined as the straight end point; is the heading angle threshold, ;
[0023] Step 2.3, any surrounding vehicles from t1 to t2 The trajectory points at the moment are marked, so as to obtain any point between t1 and t2. Labels of surrounding vehicles at time t2 ,in, Indicates the jth surrounding vehicle at time , and ,when When , it means that the jth surrounding vehicle is The behavior type at the moment is left turn, when When , it means that the jth surrounding vehicle is The behavior type at this moment is straight ahead. When , it means that the jth surrounding vehicle is The behavior type at this moment is right lane change; .
[0024] Furthermore, the step 3 includes:
[0025] Step 3.1: Calculate the jth surrounding vehicle according to formula (2) The lateral relative displacement of the vehicle ego at the moment , longitudinal relative displacement ;
[0026] (2)
[0027] Step 3.2: Use the sliding window method to perform window processing on the labeled surrounding vehicle trajectory data between time t1 and time t2 to obtain m feature subsequences. ,in, represents the i-th feature subsequence, and ,in, for The trajectory data of the main vehicle ego at the kth moment, for The trajectory data of the jth surrounding vehicle at the kth moment, for The lateral displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment is, for The longitudinal displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment; The labels of the surrounding vehicles at the kth moment are recorded as , express The label of the jth surrounding vehicle at the kth moment, , ; , is the length of each feature subsequence.
[0028] Furthermore, the step 4.1 includes:
[0029] Step 4.1.1, the bidirectional long short-term memory neural network Bi-LSTM includes: a forward LSTM layer and a reverse LSTM layer, and Process and obtain The set of hidden states of the behavior intentions of the vehicles around the middle vehicle at the kth moment ,in, express The hidden state of the behavior intention of the jth surrounding vehicle at the kth moment;
[0030] Step 4.1.2, Softmax layer pair After processing, we get The probability set of the behavior intention of the surrounding vehicles at the kth moment ,in, express The probability of the behavior intention of the jth surrounding vehicle at the kth moment.
[0031] Furthermore, the step 4.2 includes:
[0032] Step 4.2.1, the long short-term neural network LSTM and Process and obtain The hidden state set of the trajectory output of the surrounding vehicles at the kth moment ,in, express The trajectory of the jth surrounding vehicle at the kth moment outputs the hidden state;
[0033] Step 4.2.2, the channel attention includes an average pooling layer, a fully connected layer, a sigmoid activation function layer, and Process and obtain The channel attention feature sequence of the surrounding vehicles at the kth moment in time ,in, express The channel attention features of the jth surrounding vehicle at the kth moment;
[0034] Step 4.2.3, the temporal attention includes a fully connected layer, a Relu activation function layer, a normalization layer, and Process and obtain Temporal attention feature sequence of surrounding vehicles at the kth moment ,in, express Temporal attention features of the jth surrounding vehicle at the kth moment;
[0035] Step 4.2.4, multi-scale feature fusion module and Process and obtain The fusion feature sequence of the surrounding vehicles at the kth moment ,in, represents the fusion features of the jth vehicle around it at the kth moment;
[0036] Step 4.2.5, fully connected layer pair Process and obtain The predicted trajectory results of the surrounding vehicles at the kth moment ,in, Indicating The predicted trajectory result of the jth surrounding vehicle at the kth moment, and ,in, express The predicted lateral position of the jth surrounding vehicle at the kth moment, express The predicted longitudinal position of the jth surrounding vehicle at the kth moment.
[0037] Furthermore, the total loss function Loss is calculated using formula (7) in 5:
[0038]
[0039] Mode middle, express The corresponding loss function is, express Middle horizontal position, express Middle vertical position.
[0040] The electronic device of the present invention includes a memory and a processor, and is characterized in that the memory is used to store a program that supports the processor to execute the surrounding vehicle trajectory prediction method, and the processor is configured to execute the program stored in the memory.
[0041] The present invention provides a computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and the computer program executes the steps of the surrounding vehicle trajectory prediction method when the computer program is executed by a processor.
[0042] Compared with the existing technology, the beneficial technical effects of the present invention are embodied in:
[0043] 1. Traditional methods are mostly based on simple mathematical models or statistical laws, which are difficult to cope with complex and changing traffic environments. In particular, prediction accuracy and reliability are insufficient in scenarios such as multi-lane expressways and ramp entrances. This invention uses a deep learning algorithm to automatically learn complex traffic patterns and driving behavior patterns, effectively improving the accuracy and generalization of trajectory prediction.
[0044] 2. This invention not only extracts vehicle state features (such as position, speed, and acceleration), but also incorporates interactions between the vehicle and surrounding vehicles, as well as environmental features such as road sign locations. Furthermore, it constructs a driving style recognition module, a behavioral intention recognition module, and a trajectory output module, optimizing the entire process from feature extraction to trajectory prediction. This multi-dimensional feature fusion and modular design enable the model to more comprehensively understand traffic scenarios and accurately predict the trajectories of surrounding vehicles. BRIEF DESCRIPTION OF THE DRAWINGS
[0045] Figure 1 It is a scene schematic diagram of the present invention;
[0046] Figure 2 It is a schematic diagram of the trajectory marking of the present invention;
[0047] Figure 3 Schematic diagram of the network framework of the present invention. DETAILED DESCRIPTION
[0048] In this example, in the ever-advancing field of autonomous driving, the present invention breaks through the limitations of traditional technologies and carefully designs a surrounding vehicle trajectory prediction method based on a deep learning algorithm. It is applied to the main line road scenario of a one-way expressway with at least two lanes. The scenario diagram is shown in the figure below. Figure 1 As shown, this embodiment uses three lanes in the same direction, which are recorded as the first lane, the second lane, and the third lane from the inside to the outside, and an exit ramp is set on the outside of the third lane; the surrounding vehicle trajectory prediction method can accurately capture the vehicle's motion characteristics and behavioral intentions, and combine complex road environment information to achieve accurate prediction of the surrounding vehicle trajectory. Specifically, the method is carried out according to the following method.
[0049] Step 1: Use laser radar and millimeter microwave radar sensors to collect the trajectory data of the main vehicle ego and surrounding vehicles and preprocess them to obtain the preprocessed trajectory data. Let the trajectory point of the main vehicle ego at time t be ; Let the jth surrounding vehicle of the main vehicle ego at time t be recorded as , j=1,2,…,n, where T is the total sampling time, represents the lateral position of the jth surrounding vehicle at time t, represents the longitudinal position of the jth surrounding vehicle at time t, represents the speed of the jth surrounding vehicle at time t, represents the acceleration of the jth surrounding vehicle at time t; n represents the total number of surrounding vehicles, represents the lateral position of the vehicle ego at time t, represents the longitudinal position of the main vehicle ego at time t, represents the speed of the vehicle ego at time t, represents the acceleration of the vehicle ego at time t.
[0050] Step 2: Label the pre-processed trajectory data of surrounding vehicles to obtain Labels of surrounding vehicles at all times ;in, Indicates that the jth surrounding vehicle is The label of the moment, the trajectory annotation diagram is as follows Figure 2 shown.
[0051] Step 2.1: Calculate the heading angle of the jth surrounding vehicle at time t according to formula (1): ;
[0052] (1)
[0053] In formula (1), represents the lateral position of the jth surrounding vehicle at time t-1, represents the longitudinal position of the jth surrounding vehicle at time t-1;
[0054] Step 2.2: For the surrounding vehicles that have changed lanes, record the trajectory point that passes through the lane boundary at time t0 during the lane change process as the lane change point, and traverse along both sides of the lane change point:
[0055] When the heading angle value of the jth surrounding vehicle at time t1 satisfy When the trajectory point at time t1 is defined as the lane change starting point, the heading angle value at time t2 is satisfy , define the trajectory point at time t2 as the lane-changing end point;
[0056] For the surrounding vehicles that have not changed lanes, the trajectory point at time t1 is defined as the straight start point, and the trajectory point at time t2 is defined as the straight end point; where, is the heading angle threshold, .
[0057] Step 2.3, any surrounding vehicles from t1 to t2 The trajectory points at the moment are marked, so as to obtain any point between t1 and t2. Labels of surrounding vehicles at time t2 ,in, Indicates the jth surrounding vehicle at time , and ,when When , it means that the jth surrounding vehicle is The behavior type at the moment is left turn, when When , it means that the jth surrounding vehicle is The behavior type at this moment is straight ahead. When , it means that the jth surrounding vehicle is The behavior type at this moment is right lane change; .
[0058] Step 3: Extract state features of the main vehicle (ego) and surrounding vehicles, including lateral and longitudinal position features, velocity features, acceleration features, and interaction features between the main vehicle (ego) and surrounding vehicles, including the lateral and longitudinal displacements of surrounding vehicles relative to the main vehicle (ego).
[0059] Step 3.1: Calculate the jth surrounding vehicle according to formula (2) The lateral relative displacement of the vehicle ego at the moment , longitudinal relative displacement ;
[0060] (2)
[0061] Step 3.2: Use the sliding window method to perform window processing on the labeled surrounding vehicle trajectory data between time t1 and time t2 to obtain m feature subsequences. ,in, represents the i-th feature subsequence, and ,in, for The trajectory data of the main vehicle ego at the kth moment, for The trajectory data of the jth surrounding vehicle at the kth moment, for The lateral displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment is, for The longitudinal displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment; The labels of the surrounding vehicles at the kth moment are recorded as , express The label of the jth surrounding vehicle at the kth moment, , ; , is the length of each feature subsequence.
[0062] Step 4: Build a trajectory prediction network for surrounding vehicles. The network framework diagram is as follows: Figure 3 As shown, it includes a behavior intention recognition module and a trajectory output module;
[0063] Step 4.1: Construct a behavioral intention recognition module, which includes a bidirectional long short-term memory neural network Bi-LSTM and a Softmax layer. Process and obtain The probability of the behavior intention of the surrounding vehicles ;
[0064] Step 4.1.1, Bidirectional long short-term memory neural network Bi-LSTM, including: forward LSTM layer and reverse LSTM layer, and Process and obtain The set of hidden states of the behavior intentions of the vehicles around the middle vehicle at the kth moment ,in, express The hidden state of the behavior intention of the jth surrounding vehicle at the kth moment;
[0065] Step 4.1.2, Softmax layer pair After processing, we get The probability set of the behavior intention of the surrounding vehicles at the kth moment ,in, express The probability of the behavior intention of the jth surrounding vehicle at the kth moment.
[0066] Step 4.2, build the trajectory output module, including: a long short-term neural network LSTM, a channel attention layer, a time attention layer, a multi-scale feature fusion module and a fully connected layer, and and Processing is performed to obtain trajectory prediction results ;
[0067] Step 4.2.1, Long Short-Term Neural Network LSTM and Process and obtain The hidden state set of the trajectory output of the surrounding vehicles at the kth moment ,in, express The trajectory of the jth surrounding vehicle at the kth moment outputs the hidden state.
[0068] Step 4.2.2, channel attention, including an average pooling layer, a fully connected layer, a sigmoid activation function layer, and Process and obtain The channel attention feature sequence of the surrounding vehicles at the kth moment in time ,in, express The channel attention features of the jth surrounding vehicle at the kth moment;
[0069] Step 4.2.3, temporal attention, including a fully connected layer, a Relu activation function layer, a normalization layer, and Process and obtain Temporal attention feature sequence of surrounding vehicles at the kth moment ,in, express The temporal attention features of the jth surrounding vehicle at the kth moment.
[0070] Step 4.2.4, multi-scale feature fusion module and Process and obtain The fusion feature sequence of the surrounding vehicles at the kth moment ,in, represents the fusion features of the jth vehicle around it at the kth moment;
[0071] Step 4.2.5, fully connected layer pair Process and obtain The predicted trajectory results of the surrounding vehicles at the kth moment ,in, Indicating The predicted trajectory result of the jth surrounding vehicle at the kth moment, and ,in, express The predicted lateral position of the jth surrounding vehicle at the kth moment, express The predicted longitudinal position of the jth surrounding vehicle at the kth moment.
[0072] Step 5: Use formula (7) to calculate the total loss function Loss of the trajectory prediction network of surrounding vehicles;
[0073]
[0074] Mode middle, express The corresponding loss function is, express Middle horizontal position, express mid-longitudinal position;
[0075] Step 6: Use the Adam optimizer to train the trajectory prediction network of surrounding vehicles and continuously optimize the total loss function Loss to update the network parameters. When the total loss function Loss converges, the optimal surrounding vehicle trajectory prediction model is obtained, which is used to predict the trajectory of the input trajectory data.
[0076] In this embodiment, an electronic device includes a memory and a processor, wherein the memory is used to store a program that supports the processor to execute the above method, and the processor is configured to execute the program stored in the memory.
[0077] In this embodiment, a computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the above method are executed.
[0078] In summary, the method presented in this paper, by integrating advanced data acquisition techniques, deep learning algorithms, and multi-dimensional feature extraction, enables autonomous vehicles to accurately predict the trajectories of surrounding vehicles in complex traffic environments, ensuring safe and efficient operation under complex and ever-changing road conditions. The application of this technology is not limited to improving the performance of individual vehicles but is also committed to optimizing the entire transportation network. By intelligently regulating traffic flows in diversion zones, this technology can effectively alleviate traffic congestion and significantly improve road capacity.
[0079] In this embodiment, the method concept of the present invention is not limited to the trajectory prediction of the diversion flow area of a one-way three-lane expressway. Other embodiments obtained by ordinary technicians in this field without creative changes are within the scope of protection of the present invention.
Claims
1. A method for predicting surrounding vehicle trajectories based on a deep learning algorithm is applied to a one-way expressway mainline road scenario with at least two lanes, characterized in that: The surrounding vehicle trajectory prediction method is performed as follows: Step 1: Use LiDAR and millimeter-wave radar sensors to collect and preprocess the trajectory data of the main vehicle (ego) and surrounding vehicles to obtain the preprocessed trajectory data of the main vehicle (ego) and surrounding vehicles; Step 2: Mark the pre-processed trajectory data of surrounding vehicles to obtain any Labels of surrounding vehicles at all times ;in, Indicates that the jth surrounding vehicle is Labels of moments; Step 3: Use the sliding window method to perform window processing on the labeled surrounding vehicle trajectory data between time t1 and time t2 to obtain m feature subsequences ,in, represents the i-th feature subsequence; Step 4: Build a trajectory prediction network for surrounding vehicles, including a behavior intention recognition module and a trajectory output module. Step 4.1: Construct a behavioral intention recognition module, which includes a bidirectional long short-term memory neural network Bi-LSTM and a Softmax layer. Process and obtain The probability of the behavior intention of the surrounding vehicles ; Step 4.2, build the trajectory output module, including: a long short-term neural network LSTM, a channel attention layer, a time attention layer, a multi-scale feature fusion module and a fully connected layer, and and Processing is performed to obtain trajectory prediction results ; Step 5: Construct the total loss function Loss of the trajectory prediction network of surrounding vehicles; Step 6: Use the Adam optimizer to train the trajectory prediction network of surrounding vehicles and continuously optimize the total loss function Loss to update the network parameters. When the total loss function Loss converges, the optimal surrounding vehicle trajectory prediction model is obtained, which is used to predict the trajectory of the input trajectory data.
2. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 1, characterized in that: In step 1, let the trajectory point of the pre-processed main vehicle ego at time t be ; Let the trajectory point of the jth surrounding vehicle at time t after the main vehicle ego preprocessing be recorded as , j=1,2,…,n, where T is the total sampling time, represents the lateral position of the jth surrounding vehicle at time t, represents the longitudinal position of the jth surrounding vehicle at time t, represents the speed of the jth surrounding vehicle at time t, represents the acceleration of the jth surrounding vehicle at time t; n represents the total number of surrounding vehicles, represents the lateral position of the vehicle ego at time t, represents the longitudinal position of the main vehicle ego at time t, represents the speed of the vehicle ego at time t, represents the acceleration of the vehicle ego at time t.
3. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 2, characterized in that: The step 2 includes: Step 2.1: Calculate the heading angle of the jth surrounding vehicle at time t according to formula (1): ; (1) In formula (1), represents the lateral position of the jth surrounding vehicle at time t-1, represents the longitudinal position of the jth surrounding vehicle at time t-1; Step 2.2: For the surrounding vehicles that have changed lanes, record the trajectory point that passes through the lane boundary at time t0 during the lane change process as the lane change point, and traverse along both sides of the lane change point: When the heading angle value of the jth surrounding vehicle at time t1 satisfy When the trajectory point at time t1 is defined as the lane change starting point, the heading angle value at time t2 is satisfy , define the trajectory point at time t2 as the lane-changing end point; For the surrounding vehicles that have not changed lanes, the trajectory point at time t1 is defined as the straight start point, and the trajectory point at time t2 is defined as the straight end point; is the heading angle threshold, ; Step 2.3, any surrounding vehicles between t1 and t2 The trajectory points at the moment are marked, so as to obtain any point between t1 and t2. Labels of surrounding vehicles at time t2 ,in, Indicates the jth surrounding vehicle at time , and ,when When , it means that the jth surrounding vehicle is The behavior type at the moment is left turn, when When , it means that the jth surrounding vehicle is The behavior type at this moment is straight ahead. When , it means that the jth surrounding vehicle is The behavior type at this moment is right lane change; .
4. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 3, characterized in that: The step 3 includes: Step 3.1: Calculate the jth surrounding vehicle according to formula (2) The lateral relative displacement of the vehicle ego at the moment , longitudinal relative displacement ; (2) Step 3.2: Use the sliding window method to perform window processing on the labeled surrounding vehicle trajectory data between time t1 and time t2 to obtain m feature subsequences. ,in, represents the i-th feature subsequence, and ,in, for The trajectory data of the main vehicle ego at the kth moment, for The trajectory data of the jth surrounding vehicle at the kth moment, for The lateral displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment is, for The longitudinal displacement of the jth surrounding vehicle relative to the main vehicle at the kth moment; The labels of the surrounding vehicles at the kth moment are recorded as , express The label of the jth surrounding vehicle at the kth moment, , ; , is the length of each feature subsequence.
5. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 4, characterized in that: The step 4.1 includes: Step 4.1.1, the bidirectional long short-term memory neural network Bi-LSTM includes: a forward LSTM layer and a reverse LSTM layer, and Process and obtain The set of hidden states of the behavior intentions of the vehicles around the middle vehicle at the kth moment ,in, express The hidden state of the behavior intention of the jth surrounding vehicle at the kth moment; Step 4.1.2, Softmax layer pair After processing, we get The probability set of the behavior intention of the surrounding vehicles at the kth moment ,in, express The probability of the behavior intention of the jth surrounding vehicle at the kth moment.
6. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 5, characterized in that: The step 4.2 includes: Step 4.2.1, the long short-term neural network LSTM and Process and obtain The hidden state set of the trajectory output of the surrounding vehicles at the kth moment ,in, express The trajectory of the jth surrounding vehicle at the kth moment outputs the hidden state; Step 4.2.2, the channel attention includes an average pooling layer, a fully connected layer, a sigmoid activation function layer, and Process and obtain The channel attention feature sequence of the surrounding vehicles at the kth moment in time ,in, express The channel attention features of the jth surrounding vehicle at the kth moment; Step 4.2.3, the temporal attention includes a fully connected layer, a Relu activation function layer, a normalization layer, and Process and obtain Temporal attention feature sequence of surrounding vehicles at the kth moment ,in, express Temporal attention features of the jth surrounding vehicle at the kth moment; Step 4.2.4, multi-scale feature fusion module and Process and obtain The fusion feature sequence of the surrounding vehicles at the kth moment ,in, represents the fusion features of the jth vehicle around it at the kth moment; Step 4.2.5, fully connected layer pair Process and obtain The predicted trajectory results of the surrounding vehicles at the kth moment ,in, Indicating The predicted trajectory result of the jth surrounding vehicle at the kth moment, and ,in, express The predicted lateral position of the jth surrounding vehicle at the kth moment, express The predicted longitudinal position of the jth surrounding vehicle at the kth moment.
7. The method for predicting surrounding vehicle trajectories based on a deep learning algorithm according to claim 6, characterized in that: In the above 5, the total loss function Loss is calculated using formula (7): (7) Mode middle, express The corresponding loss function is, express Middle horizontal position, express Middle vertical position.
8. An electronic device comprising a memory and a processor, characterized in that: The memory is used to store a program that supports the processor to execute the surrounding vehicle trajectory prediction method according to any one of claims 1 to 7, and the processor is configured to execute the program stored in the memory.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the surrounding vehicle trajectory prediction method according to any one of claims 1 to 7 are executed.