Trajectory prediction method based on radar multi-target tracking

By combining AIMM-UKF-JPDA algorithm and intention-space-time dynamic attention network in multi-objective tracking and trajectory prediction, the accuracy and efficiency of multi-objective tracking and trajectory prediction in complex traffic environments is solved, and higher tracking is achieved Accuracy and trajectory prediction reliability enhance the safety of autonomous vehicles.

CN118915071APending Publication Date: 2024-11-08GUILIN UNIV OF ELECTRONIC TECH
View PDF 0 Cites 8 Cited by

Patent Information

Application Number
CN202411028959.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-30
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

Existing multi-objective tracking and trajectory prediction technologies face the dual challenges of accuracy and efficiency in complex traffic environments, making it difficult to provide accurate vehicle status and future driving path predictions in real-time processing.

Method used

Using a radar-based multi-objective tracking method, combined with AIMM-UKF-JPDA algorithm and intention-space-time dynamic attention trajectory prediction network, the timing features are extracted by allocating tracking targets for different survival cycles, using improved filters and data association algorithms, and extracting timing features. Social interaction and space-time conversion are processed through BiLSTM networks, and ultimately predict the future intentions and trajectories of the vehicle through the decoder.

Benefits of technology

It improves the accuracy of multi-objective tracking and the reliability of trajectory prediction, provides more accurate vehicle status information and future driving trajectory, and enhances the safety and adaptability of autonomous vehicles in complex environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118915071A_ABST
    Figure CN118915071A_ABST
Patent Text Reader

Abstract

According to the trajectory prediction method based on radar multi-target tracking, the improved AIMM-UKF-JPDA algorithm is adopted for vehicle tracking, and the trajectory is predicted in combination with the intention-space-time dynamic attention network. In the aspect of multi-target tracking, the system divides tracking targets into a high-confidence-coefficient type and a low-confidence-coefficient type, and data association is carried out by using a JPDA algorithm and an NN algorithm respectively, so that the tracking precision is improved. And an association result is fed back to the AIMM algorithm, and the target state and the probability matrix are updated. And associating the unassociated target with the remaining targets at the previous moment to generate a new tracking target. In the aspect of trajectory prediction, the system inputs 70% of tracking data to an encoder for feature extraction, time sequence features are encoded through a BiLSTM network, and social interaction and time-space conversion are processed by using a multi-head attention mechanism. The decoder predicts the vehicle intention and the future trajectory through LSTM decoding in combination with the context information of the encoder and the intention recognition result. According to the method, MSE (root-mean-square error) is adopted for performance evaluation, so that the accuracy and reliability of trajectory prediction are ensured.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle multi-target tracking and trajectory prediction, and in particular to an AIMM-UKF-JPDA vehicle multi-target tracking and an intention-space-time dynamic attention trajectory prediction network. Background Art

[0002] With the rapid development of intelligent transportation systems and autonomous driving technology, multi-target tracking and trajectory prediction technology are particularly important. Multi-target tracking is the basis for understanding the dynamics of each vehicle in a traffic scene. It can monitor and record the position, speed, acceleration and other information of the vehicle in real time. The implementation of this technology has a significant effect on improving road safety, optimizing traffic flow management, and reducing the incidence of traffic accidents. Especially in urban traffic environments, multi-target tracking can effectively identify and distinguish individual vehicles in dense traffic, providing accurate data support for traffic monitoring and vehicle management.

[0003] Trajectory prediction technology further utilizes these tracking data to predict the possible future driving path of the vehicle by analyzing its historical movement pattern and current state. This is crucial for autonomous vehicles when making path planning and decisions. Accurate trajectory prediction can significantly improve the autonomy and adaptability of autonomous vehicles, enabling them to make fast and safe driving decisions in complex traffic environments. In addition, trajectory prediction also has a wide range of application value in areas such as vehicle navigation systems, traffic signal control, and traffic congestion analysis.

[0004] However, due to the complexity of the traffic environment, the uncertainty of vehicle behavior and the high requirement of real-time processing, existing multi-target tracking and trajectory prediction technologies face the dual challenges of accuracy and efficiency. Summary of the invention

[0005] The purpose of the present invention is to provide a trajectory prediction method based on radar multi-target tracking in order to overcome the above problems and improve the performance of tracking and prediction.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] A trajectory prediction method based on radar multi-target tracking includes the following steps:

[0008] Step 1: According to the tracking duration, the vehicle tracking targets are divided into two categories: high confidence and low confidence, and different life cycles are assigned to each category of targets to optimize the tracking accuracy.

[0009] Step 2: Use the improved AIMM algorithm to model the vehicle's motion state and calculate the target's interaction state estimate based on the model's state estimate and probability.

[0010] Step 3: Apply the AIMM-UKF filter to predict the mixed measurement state and its covariance of the target.

[0011] Step 4: Perform data association, using the JPDA algorithm for high-confidence targets and the NN algorithm for low-confidence targets.

[0012] Step 5: Update the target state and the probability transfer matrix of the AIMM algorithm based on the data association results.

[0013] Step 6: Perform data association on the unsuccessfully associated detection targets to generate new tracking targets.

[0014] Step 7: Output the features of the correctly tracked target and preprocess it, dividing it into a 7:2:1 dataset, where 70% is the training set, 20% is the validation set, and 10% is the test set.

[0015] Step 8: Use the BiLSTM network to extract temporal features from the historical trajectory data of the vehicle, use the multi-head attention mechanism to process social interactions and spatiotemporal transformations, and output contextual information.

[0016] Step 9: The decoder recognizes the intention of the current step based on the context information, decodes the intention recognition and context information together through LSTM, and predicts the vehicle intention and future trajectory of the current step.

[0017] Step 10: Use the MSE mean square error function, NLL likelihood loss function, etc. to train and save the model. After training, the historical trajectory of multi-target tracking and identification can be input into the trajectory prediction model, and the model outputs future intentions and trajectories.

[0018] Preferably, the step 1 specifically includes:

[0019] S1.1: Track the associated frame count of each target. In each frame of point cloud data, if the target is successfully associated, the count is increased by one, and if it is not associated, it is decreased by one. The tracking time is determined by this count.

[0020] S1.2: Set a threshold and classify the targets into high confidence and low confidence categories based on the comparison between the tracking time and the threshold.

[0021] S1.3: Set a life cycle for each type of target. If a target is not associated with a certain number of frames for a period exceeding its life cycle, it will be marked as to be eliminated.

[0022] Preferably, the step 2 specifically includes:

[0023] Through the improved Adaptive Importance Monte Carlo (AIMM) algorithm, the system models the motion state of the vehicle target and uses three models, namely, uniform straight line (CV), uniform acceleration (CA), and uniform turning rate and speed (CTRV), to simulate the actual driving behavior of the vehicle. Initially, each model is assigned the same probability, i.e., 1 / 3. The system uses the state estimates and probabilities of these models to dynamically update the posterior probability of the target through the probability transfer matrix, thereby calculating the interactive state estimates of the target under different motion states.

[0024] Preferably, the step 3 specifically includes:

[0025] S3.1: Calculate sigma points

[0026]

[0027] S3.2: Compute a priori state estimates and covariance

[0028]

[0029] S3.3: Compute measurement state estimates and covariance

[0030]

[0031] S3.4: Calculate the mixed measurement state estimate z k / k-1 and covariance S k / k-1

[0032]

[0033] Preferably, the step 4 specifically includes:

[0034] S4.1: The system evaluates the distances between high-confidence tracking targets and each detected target, and only retains those targets whose distances are less than a preset threshold to form a correlation matrix 1. The remaining detected targets whose distances exceed the threshold and the unassociated tracking targets are temporarily stored separately for subsequent processing.

[0035] S4.2: The system performs a clustering operation on the association matrix 1, classifying the interconnected tracking targets and detection targets into the same group, thereby dividing the entire matrix into multiple sub-matrices, each of which independently reflects the association status of a group of targets.

[0036] S4.3: Apply the JPDA algorithm to each sub-association matrix to determine the posterior probability of each tracked target and its corresponding detected target, and weighted average the detected targets accordingly to update the state estimate of the tracked target.

[0037] S4.4: Perform detection target pruning, calculate the distance between the updated tracking target and the detection target, and generate a weighted association matrix based on it. Use the Hungarian algorithm to find the minimum distance match, and mark the detection and tracking targets that do not meet the distance threshold as pending association.

[0038] S4.5: For low confidence targets, the system calculates the distance between them and the target to be associated 1, creates a weighted association matrix, and uses the Hungarian algorithm to determine the optimal match. Objects that do not meet the distance threshold are reclassified as target to be associated 2 and unassociated tracking target 1.

[0039] Preferably, the step 5 specifically includes:

[0040] S5.1: Check the tracking time of the unassociated tracking target 1 in step 4, eliminate the targets that exceed the life cycle, and use the prior estimation of the remaining ones as the current optimal state.

[0041] S5.2: In step 4, the matched detected target is taken as the true measurement, and the probability and state of all tracking target sub-models are updated through maximum likelihood estimation and UKF filter. The high confidence target has been updated through the JPDA algorithm.

[0042] S5.3: Using the updated model probability, perform a weighted average of the optimal estimates of each tracking target to obtain a comprehensive state estimate.

[0043] S5.4: Update the probability transfer matrix of the tracking target according to the rate of change of the model probability at adjacent moments as a correction parameter.

[0044] S5.5: Update the tracking frame information of the associated tracking target according to the detection frame of the detection target. For the target that is not associated but the tracking time is less than the life cycle, the tracking frame information of the previous moment is continued.

[0045] Preferably, the step 6 specifically includes:

[0046] After step 4, the detection target 2 to be associated is matched with the remaining detection targets at the previous moment. The correlation matrix is ​​constructed by calculating the distance and IOU (intersection over union), and the Hungarian algorithm is applied to find the minimum cost match. Those detection targets that meet the threshold conditions in distance and IOU are selected to initialize the new tracking target, and the remaining detection targets that do not meet the conditions are used as the remaining detection targets at the current moment.

[0047] Preferably, the step 7 specifically includes:

[0048] S7.1: Collect the data output by the multi-target tracking module and collect the speed, acceleration, x, y and other characteristic information of all vehicles.

[0049] S7.2: Perform data preprocessing operations such as filtering and smoothing on the output data.

[0050] S7.3: When the heading angle exceeds the threshold three times in a row, the lateral intention is generated, and the labels are set to go straight (1), change lanes left (2), and change lanes right (3). When the future acceleration is greater than the threshold, the longitudinal intention is generated, and the labels are set to uniform speed (1), acceleration (2), and deceleration (3).

[0051] S7.4: Divide the dataset into training, validation, and test sets in a ratio of 7:2:1.

[0052] S7.5: Divide the map into 13*3 grids with the predicted vehicle as the center, and fill in the positions of all vehicles.

[0053] S7.6: Perform sliding window sampling on the processed data, selecting 3s as the historical trajectory and 5s as the future prediction trajectory.

[0054] Preferably, the step 8 specifically includes:

[0055] S8.1: Time series extraction module: The agent state information is converted into an embedding vector through a linear layer and then input into a bidirectional LSTM (BI-LSTM) for encoding.

[0056]

[0057] Among them, x t is the activated embedding vector h t-1 is the hidden state of the previous time step, h t is the hidden state of the current step output, and W,b is the update parameter. ...represents the hidden state of the predicted agent at each time stamp, and ...represents the hidden state of vehicle i at each timestamp.

[0058] S8.2: Social interaction module: First, a linear layer is used to convert the hidden state of the historical trajectory of the predicted agent and the hidden state of the historical trajectory of the surrounding agents into three different vectors. Then three learnable linear transformations are used to convert the features of the target vehicle and the surrounding vehicles into query, key, and value vectors. The dot product attention mechanism is used to calculate the attention weights between the target vehicle and each surrounding vehicle. These weights reflect the degree of influence of the surrounding vehicles on the target vehicle at a specific point in time. According to the calculated attention weights, the features of the surrounding vehicles are weighted summed, and this information is aggregated to the target vehicle, thereby updating the state of the target vehicle. Finally, for each timestamp, the module will output an updated feature sequence that reflects the social interaction between the target vehicle and the surrounding vehicles.

[0059]

[0060] S8.3: Spatiotemporal interaction module: It receives the output sequence of the social interaction module and uses a multi-head self-attention mechanism to process time series data. For each timestamp, a learnable parameter matrix is ​​used to convert the input sequence into a query, key, and value matrix. The attention score matrix is ​​calculated by the dot product operation of the query and the key, which represents the degree of mutual influence between different timestamps. According to the attention score, the value matrix is ​​weighted and summed to aggregate the information across timestamps to obtain a feature representation that integrates time dependencies. Finally, the spatiotemporal interaction module will generate an updated feature sequence that characterizes the time dependencies of social interactions at different timestamps, providing further time context information for trajectory prediction.

[0061]

[0062] Preferably, the step 9 specifically includes:

[0063] S9.1: The decoder receives the context information output by the encoder and uses a fully connected layer and a softmax activation function to identify the possible driving intentions of the target vehicle, such as keeping lanes, changing lanes, or accelerating, and calculates the probabilities of these intentions. For each possible driving intention, the IFF module assigns weights to features at different timestamps. These weights reflect the contribution of historical trajectory features to predicting future trajectories under a specific intention.

[0064] According to the assigned weights, the features of different timestamps are weightedly fused.

[0065]

[0066] Among them, L lat For left lane change, right lane change and going straight, Llon For acceleration, constant speed and deceleration.

[0067] S9.2: The decoder concatenates the horizontal and vertical intentions predicted by softmax into the context information and uses all of them as the input of the LSTM network. The output of LSTM is the obtained predicted distribution parameters.

[0068]

[0069] θ t′ =MLP(h t′ ,W pre )

[0070] in, is the fusion vector of horizontal and vertical intention and context information, θ t′ is the prediction distribution parameter, W LSTM is the weight of the LSTM unit, W pre is the hidden state h t′-1 The weights used to convert to bivariate Gaussian distribution parameters.

[0071] Preferably, the step 10 specifically includes:

[0072] S10.1: The predicted parameter distribution output by the decoder is calculated through the MSE mean square error loss function and the NLL likelihood loss function. The parameters in the back propagation error learning model are used to minimize the loss. The MSE loss function calculation formula is as follows:

[0073]

[0074] Among them, (xt, yt) is the actual predicted position of the vehicle at timestamp t, and μt, x, μt, y is the position predicted by the model.

[0075] The NLL loss function calculation formula is as follows:

[0076]

[0077] S10.2: During the training process, we set the epoch to 20 times. The epoch with the highest validation set accuracy and no overfitting phenomenon among the 20 times is taken as the optimal model, and the model is saved.

[0078] S10.3: Input the test set data into the saved model, and the output is the future trajectory and intention of the vehicle.

[0079] The beneficial effects of the present invention are:

[0080] The present invention provides a trajectory prediction method based on radar multi-target tracking, and designs for the first time a vehicle multi-target tracking based on AIMM-UKF-JPDA and a trajectory prediction network framework based on intention-space-time dynamic attention. By using a high-precision multi-target tracking algorithm, more accurate vehicle status information is obtained, providing more reliable data support for the downstream trajectory prediction network. Trajectory prediction uses the intention-space-time dynamic attention network to input the precise information of the upstream module. The encoder captures valuable information of the vehicle through the timing module, dynamic social interaction module and space-time module; the decoder adopts a unique intention recognition module to decode the future driving trajectory and intention output of the vehicle through the LSTM network. Through more accurate multi-target tracking and trajectory prediction, more surrounding vehicle information is provided to the downstream path planning and decision-making modules, which can effectively improve the safety and reliability of autonomous driving in a dynamic environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0082] Figure 1 It is a flow chart of a trajectory prediction method based on radar multi-target tracking of the present invention.

[0083] Figure 2 This is the overall framework diagram of the trajectory prediction model.

[0084] Figure 3 It is a schematic diagram of the horizontal and vertical intention mode. DETAILED DESCRIPTION

[0085] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with examples. It should be understood that the specific implementation described herein is only used to explain the present invention, and is not used to limit the present invention.

[0086] The following are some nouns and terminology descriptions of the present invention:

[0087] IMM: Interacting Multiple Model, interactive multiple model;

[0088] UKF: Unscented Kalman Filter, unscented Kalman filter;

[0089] AIMM: Adaptive IMM, adaptive interactive multi-model;

[0090] JPDA: Joint Probabilistic Data Association, joint probabilistic data association;

[0091] NN: NearestNeighbor, nearest neighbor data association.

[0092] This example provides a method for multi-target tracking and trajectory prediction of surrounding vehicles by using an unmanned vehicle equipped with multi-line laser radar in urban traffic. In this solution, the target detection algorithm is first used as the data source for AIMM-UKF-JPDA multi-target tracking, and the tracking module outputs the feature information of the surrounding vehicle. Then the trajectory prediction module accesses the output information and predicts the future intention and trajectory of the surrounding vehicle through an encoder-decoder model, providing more information for downstream modules.

[0093] See also Figure 1 , which is a flowchart of the entire method, including the following steps:

[0094] S1: The confidence level of the vehicle tracking target is divided according to the tracking time, and different life cycles are set for tracking targets with different confidence levels;

[0095] S2: The AIMM algorithm models the vehicle target motion state and calculates the interactive state estimate of the tracked target;

[0096] S3: AIMM-UKF filter predicts the interactive measurement state estimate and covariance of the tracked target;

[0097] S4: Data association is performed between the prediction results and the detected targets, and different data association algorithms are used for tracking targets with different confidence levels;

[0098] S5: Update the state of the tracking target and the probability transfer matrix of the AIMM algorithm according to the result of data association;

[0099] S6: Data association is performed between the unsuccessfully associated detection targets and the remaining detection targets at the previous moment to generate new tracking targets.

[0100] S7: Output the features of the correctly tracked target and preprocess it, dividing it into a 7:2:1 data set, where 70% is the training set, 20% is the validation set, and 10% is the test set.

[0101] S8: The historical trajectory data of the vehicle is extracted through the BiLSTM network to extract temporal features, and the multi-head attention mechanism is used to process social interactions and spatiotemporal transformations to output contextual information.

[0102] S9: The decoder recognizes the intention of the current step based on the context information, decodes the intention recognition and context information together through LSTM, and predicts the vehicle intention and future trajectory of the current step.

[0103] S10: The model is trained and saved using the MSE mean square error function, the NLL likelihood loss function, etc. After the training is completed, the historical trajectory of multi-target tracking and identification can be input into the trajectory prediction model, and the model outputs the future intention and trajectory.

Claims

1. A trajectory prediction method based on radar multi-target tracking, characterized in that The following steps are involved: Step 1: According to the tracking duration, the vehicle tracking targets are divided into two categories: high confidence and low confidence, and different life cycles are assigned to each category of targets to optimize the tracking accuracy. Step 2: Use the improved AIMM algorithm to model the vehicle's motion state and calculate the target's interaction state estimate based on the model's state estimate and probability. Step 3: Apply the AIMM-UKF filter to predict the mixed measurement state and its covariance of the target. Step 4: Perform data association, using the JPDA algorithm for high-confidence targets and the NN algorithm for low-confidence targets. Step 5: Update the target state and the probability transfer matrix of the AIMM algorithm based on the data association results. Step 6: Perform data association on the unsuccessfully associated detection targets to generate new tracking targets. Step 7: Output the features of the correctly tracked target and preprocess it, dividing it into a 7:2:1 dataset, where 70% is the training set, 20% is the validation set, and 10% is the test set. Step 8: Use the BiLSTM network to extract temporal features from the historical trajectory data of the vehicle, use the multi-head attention mechanism to process social interactions and spatiotemporal transformations, and output contextual information. Step 9: The decoder recognizes the intention of the current step based on the context information, decodes the intention recognition and context information together through LSTM, and predicts the vehicle intention and future trajectory of the current step. Step 10: Use the MSE mean square error function, NLL likelihood loss function, etc. to train and save the model. After training, the historical trajectory of multi-target tracking and identification can be input into the trajectory prediction model, and the model outputs future intentions and trajectories.

2. A trajectory prediction method based on radar multi-target tracking, wherein step 1 is characterized in that: S1.1: Track the associated frame count of each target. In each frame of point cloud data, if the target is successfully associated, the count is increased by one, and if it is not associated, it is decreased by one. The tracking time is determined by this count. S1.2: Set a threshold and classify the targets into high confidence and low confidence categories based on the comparison between the tracking time and the threshold. S1.3: Set a life cycle for each type of target. If a target is not associated with a certain number of frames for a period exceeding its life cycle, it will be marked as to be eliminated.

3. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 2 is characterized in that: Through the improved Adaptive Importance Monte Carlo (AIMM) algorithm, this system models the motion state of the vehicle target and uses three models, namely, uniform straight line (CV), uniform acceleration (CA), and uniform turning rate and speed (CTRV), to simulate the actual driving behavior of the vehicle. Initially, each model is assigned the same probability, i.e., 1 / 3. The system uses the state estimates and probabilities of these models to dynamically update the posterior probability of the target through the probability transfer matrix, thereby calculating the interactive state estimate of the target under different motion states. The specific expression is as follows: in represents the probability that the target is in model j after the input interaction, π j|i To represent the probability of model i transferring to model j in the probability transfer matrix, is the probability that the model is i before the interaction; represents the probability that the target is interacted by model i, assuming that the target is in model j after interaction; and Represents state estimates and covariances for different models; and Represents the interaction state estimates and interaction covariances of the UKF filter inputs under different models.

4. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 3 is characterized in that: Using the AIMM-UKF filter, this system accurately predicts the future state of the vehicle tracking target. This process first uses the Unscented Kalman Filter (UKF) to independently predict the state and uncertainty (covariance) of each motion model at the next moment. Then, based on the probability of each model, a weighted average method is used to synthesize a comprehensive state prediction and covariance to form a hybrid measurement estimate of the target's future position and velocity. It includes the following steps: S3.1: Calculation of sigma points S3.2: Compute a priori state estimates and covariance S3.3: Compute measurement state estimates and covariance S3.4: Calculate the mixed measurement state estimate z k / k-1 and covariance S k / k-1 5. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 4 is characterized in that: The prediction results are data associated with the detected targets. The JPDA data association algorithm is used for high-confidence tracking targets, and the NN data association algorithm is used for low-confidence tracking targets. The specific steps of the data association phase include the following: S4.1: The system evaluates the distances between high-confidence tracking targets and each detected target, and only retains those targets whose distances are less than a preset threshold to form a correlation matrix 1. The remaining detected targets whose distances exceed the threshold and the unassociated tracking targets are temporarily stored separately for subsequent processing. S4.2: The system performs a clustering operation on the association matrix 1, classifying the interconnected tracking targets and detection targets into the same group, thereby dividing the entire matrix into multiple sub-matrices, each of which independently reflects the association status of a group of targets. S4.3: Apply the JPDA algorithm to each sub-association matrix to determine the posterior probability of each tracked target and its corresponding detected target, and weighted average the detected targets accordingly to update the state estimate of the tracked target. S4.4: Perform detection target pruning, calculate the distance between the updated tracking target and the detection target, and generate a weighted association matrix based on it. Use the Hungarian algorithm to find the minimum distance match, and mark the detection and tracking targets that do not meet the distance threshold as pending association. 6.S4.5: For low confidence targets, the system calculates the distance between them and the target to be associated 1, creates a weighted association matrix, and uses the Hungarian algorithm to determine the best match. Objects that do not meet the distance threshold are reclassified as target to be associated 2 and unassociated tracking target 1.

7. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 5 is characterized in that: The following steps are involved: S5.1: Check the tracking time of the unassociated tracking target 1 in step 4, eliminate the targets that exceed the life cycle, and use the prior estimation of the remaining ones as the current optimal state. S5.2: In step 4, the matched detected target is taken as the true measurement, and the probability and state of all tracking target sub-models are updated through maximum likelihood estimation and UKF filter. The high confidence target has been updated through the JPDA algorithm. S5.3: Using the updated model probability, perform a weighted average of the optimal estimates of each tracking target to obtain a comprehensive state estimate. S5.4: Update the probability transfer matrix of the tracking target according to the rate of change of the model probability at adjacent moments as a correction parameter. S5.5: Update the tracking frame information of the associated tracking target according to the detection frame of the detection target. For the target that is not associated but the tracking time is less than the life cycle, the tracking frame information of the previous moment is continued.

8. The trajectory prediction method based on radar multi-target tracking as claimed in claim 1, wherein step 6 is characterized in that: After step 4, the detection target 2 to be associated is matched with the remaining detection targets at the previous moment. The correlation matrix is ​​constructed by calculating the distance and IOU (intersection over union), and the Hungarian algorithm is applied to find the minimum cost match. Those detection targets that meet the threshold conditions in distance and IOU are selected to initialize the new tracking target, and the remaining detection targets that do not meet the conditions are used as the remaining detection targets at the current moment.

9. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 7 is characterized in that: S7.1: Collect the data output by the multi-target tracking module and collect the speed, acceleration, x, y and other characteristic information of all vehicles. S7.2: Perform data preprocessing operations such as filtering and smoothing on the output data. S7.3: When the heading angle exceeds the threshold three times in a row, the lateral intention is generated, and the labels are set to go straight (1), change lanes left (2), and change lanes right (3). When the future acceleration is greater than the threshold, the longitudinal intention is generated, and the labels are set to uniform speed (1), acceleration (2), and deceleration (3). S7.4: Divide the dataset into training, validation, and test sets in a ratio of 7:2:

1. S7.5: Divide the map into 13*3 grids with the predicted vehicle as the center, and fill in the positions of all vehicles. S7.6: Perform sliding window sampling on the processed data, selecting 3s as the historical trajectory and 5s as the future prediction trajectory.

10. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 8 is characterized in that: S8.1: Time series extraction module: The agent state information is converted into an embedding vector through a linear layer and then input into a bidirectional LSTM (BI-LSTM) for encoding. Among them, x t is the activated embedding vector h t-1 is the hidden state of the previous time step, h t is the hidden state of the current step output, and W,b is the update parameter. ...represents the hidden state of the predicted agent at each time stamp, and ...represents the hidden state of vehicle i at each timestamp. S8.2: Social interaction module: First, a linear layer is used to convert the hidden state of the historical trajectory of the predicted agent and the hidden state of the historical trajectory of the surrounding agents into three different vectors. Then three learnable linear transformations are used to convert the features of the target vehicle and the surrounding vehicles into query, key, and value vectors. The dot product attention mechanism is used to calculate the attention weights between the target vehicle and each surrounding vehicle. These weights reflect the degree of influence of the surrounding vehicles on the target vehicle at a specific point in time. According to the calculated attention weights, the features of the surrounding vehicles are weighted summed, and this information is aggregated to the target vehicle, thereby updating the state of the target vehicle. Finally, for each timestamp, the module will output an updated feature sequence that reflects the social interaction between the target vehicle and the surrounding vehicles. S8.3: Spatiotemporal interaction module: It receives the output sequence of the social interaction module and uses a multi-head self-attention mechanism to process time series data. For each timestamp, a learnable parameter matrix is ​​used to convert the input sequence into a query, key, and value matrix. The attention score matrix is ​​calculated by the dot product operation of the query and the key, which represents the degree of mutual influence between different timestamps. According to the attention score, the value matrix is ​​weighted and summed to aggregate the information across timestamps to obtain a feature representation that integrates time dependencies. Finally, the spatiotemporal interaction module will generate an updated feature sequence that characterizes the time dependencies of social interactions at different timestamps, providing further time context information for trajectory prediction.

11. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 9 is characterized in that: S9.1: The decoder receives the context information output by the encoder and uses a fully connected layer and a softmax activation function to identify possible driving intentions of the target vehicle, such as keeping lanes, changing lanes, or accelerating, and calculates the probabilities of these intentions. For each possible driving intention, the IFF module assigns weights to features at different timestamps. These weights reflect the contribution of historical trajectory features to predicting future trajectories under a specific intention. Features at different timestamps are weighted and fused according to the assigned weights. Among them, L lat For left lane change, right lane change and going straight, L lon For acceleration, constant speed and deceleration. S9.2: The decoder concatenates the horizontal and vertical intentions predicted by softmax into the context information and uses all of them as the input of the LSTM network. The output of LSTM is the obtained predicted distribution parameters. θ t′ =MLP(h t′ ,IN pre ) in, is the fusion vector of horizontal and vertical intention and context information, θ t′ is the prediction distribution parameter, W LSTM is the weight of the LSTM unit, W pre is the hidden state h t′-1 The weights used to convert to bivariate Gaussian distribution parameters.

12. The trajectory prediction method based on radar multi-target tracking according to claim 1, wherein step 10 is characterized in that: S10.1: The predicted parameter distribution output by the decoder is calculated through the MSE mean square error loss function and the NLL likelihood loss function. The parameters in the back propagation error learning model are used to minimize the loss. The MSE loss function calculation formula is as follows: Among them, (xt, yt) is the actual predicted position of the vehicle at timestamp t, and μt, x, μt, y is the position predicted by the model. The NLL loss function calculation formula is as follows: S10.2: During the training process, we set the epoch to 20 times. The epoch with the highest validation set accuracy and no overfitting phenomenon among the 20 times is taken as the optimal model, and the model is saved. S10.3: Input the test set data into the saved model, and the output is the future trajectory and intention of the vehicle.

Citation Information

Cited By

  • Ship target tracking method based on radar imaging in shielding environment

    CN119439147A

  • Flow and trajectory combined urban spatio-temporal data dynamic representation learning model and method

    CN119763321A

  • Vehicle trajectory prediction method and system, computer equipment and medium

    CN120279712A

  • Dynamic multi-target tracking and trajectory prediction method

    CN120355754A

  • A dynamic multi-target tracking and trajectory prediction method

    CN120355754B