A vehicle trajectory defective data repair method for roadside perception unit

Through the graph neural network and attention mechanism combined with high-precision lane-level map data, the accuracy and stability problems in the repair of vehicle trajectory data across the region are solved, and the accurate repair of vehicle trajectory defect data in different traffic scenarios is achieved, which improves the improvement of Data quality and road operation efficiency.

CN113963334BActive Publication Date: 2024-11-12TONGJI UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202111427550.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-11-29
Publication Date
2024-11-12
Estimated Expiration
2041-11-29

AI Technical Summary

Technical Problem

The prior art fails to effectively consider the movement behavior of vehicles close to each other in the repair of vehicle trajectory data in the whole-domain vehicle, resulting in insufficient repair accuracy and stability, which cannot meet the data quality requirements of applications such as autonomous driving and vehicle-road collaboration.

Method used

Using a vehicle trajectory defect data repair method based on graph neural network and attention mechanism, by building a TrajGAT network, we learn the lane graph topological relationship and the mutual influence between vehicles, and combine high-precision lane-level map data to realize vehicle trajectory defect data in the whole domain Accurate repair.

Benefits of technology

It improves the accuracy and stability of vehicle trajectory repair, can effectively repair vehicle defective motion trajectory in different traffic scenarios, provide high-quality data support, and improve the perception, operation efficiency and safety of road traffic conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113963334B_ABST
    Figure CN113963334B_ABST
Patent Text Reader

Abstract

The present invention relates to a vehicle trajectory missing data repair method for a roadside perception unit, comprising: collecting global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios to merge as a data set, dividing the data set into a training set, a validation set and a test set by preprocessing the data in the data set; constructing a vehicle trajectory missing data repair model based on a graph neural network and an attention mechanism; iteratively training the vehicle trajectory missing data repair model using the training set, and testing the repair effect of the vehicle trajectory missing data repair model using the validation set to obtain a vehicle trajectory missing data repair model with optimal parameters; inputting the test set or the vehicle trajectory data to be repaired into the vehicle trajectory missing data repair model with optimal parameters, and outputting the corresponding repaired global vehicle trajectory complete data. Compared with the prior art, the present invention can accurately and stably repair vehicle trajectory missing data in different scenarios.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of traffic information processing, and in particular to a method for repairing vehicle trajectory defective data for a roadside perception unit. Background Art

[0002] In recent years, with the development and widespread deployment of emerging roadside sensors such as high-definition cameras, millimeter-wave radars, and lidars, it is possible to effectively collect vehicle trajectory data for the entire road section. In the fields of autonomous driving and vehicle-road collaboration, these global vehicle trajectory data contain huge value and can assist in the realization of intelligent control of traffic flow in all time and space, active vehicle safety control, etc. However, due to factors such as the performance of the collection equipment, target detection and tracking algorithms, and physical occlusion, these data often have some vehicle trajectory defects, which makes it difficult for the collected data to be directly used by subsequent tasks, and the data value is greatly reduced. Therefore, in order to improve data quality and data usage value, it is necessary to repair the defective vehicle trajectory, and the processed complete vehicle trajectory will help tasks such as driving decision assistance and safety risk control.

[0003] At present, there are few methods for repairing vehicle defective trajectories. Time series-based defective trajectory repair is the most widely used method. This type of method uses multiple complete trajectory data of a target to learn the trajectory features of the randomly erased segments in the trajectory data, thereby being able to restore more defective trajectory data. However, there is no relevant research on global vehicle trajectory repair in the existing methods.

[0004] Time series-based defective trajectory repair can be applied to time series data repair tasks in many fields. Since this repair method only represents the position information in the time dimension, it does not consider the movement behavior of surrounding vehicles that are close to each other in space and have a mutual constraint relationship. As a result, its accuracy in applying to global vehicle motion trajectory data needs to be improved. On the other hand, this type of repair method only relies on the trajectory of a single target itself for repair. Affected by data noise, the repair result fluctuates greatly and the stability of the repair prediction data cannot be guaranteed. Summary of the invention

[0005] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and provide a vehicle trajectory missing data repair method for a roadside perception unit, so as to accurately and stably repair the global vehicle trajectory missing data.

[0006] The object of the present invention can be achieved by the following technical solution: A method for repairing vehicle trajectory defective data for a roadside perception unit, comprising the following steps:

[0007] S1. Collect global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios to merge them into a data set, and divide the data set into a training set, a validation set, and a test set by preprocessing the data in the data set;

[0008] S2. Based on graph neural network and attention mechanism, a vehicle trajectory missing data repair model is constructed;

[0009] S3, iteratively training the vehicle trajectory defective data repair model using the training set, and verifying the repair effect of the vehicle trajectory defective data repair model using the verification set, so as to obtain the vehicle trajectory defective data repair model with optimal parameters;

[0010] S4. Input the test set or the vehicle trajectory data to be repaired into the vehicle trajectory defective data repair model with optimal parameters, and output the corresponding repaired global vehicle trajectory complete data.

[0011] Furthermore, the global vehicle trajectory data is specifically a set of spatiotemporal motion trajectories composed of the motion trajectory of each individual vehicle when multiple individual vehicles move simultaneously within a certain range.

[0012] Furthermore, the step S1 specifically includes the following steps:

[0013] S11, selecting different road sections, and taking each road section as a scene;

[0014] S12, collecting global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios, and merging them as a data set;

[0015] S13. The data corresponding to the scene with missing vehicle trajectories in the data set is used as a test set, and the data corresponding to the remaining scenes are preprocessed to obtain a training set and a verification set.

[0016] Furthermore, the step S12 specifically includes the following steps:

[0017] S121, collecting, through a roadside sensing unit, trajectory data of all vehicles passing through a road section within a set time period and a lane-level map of the road section;

[0018] The collected vehicle trajectory data and lane-level map are converted into ground-based two-dimensional bird's-eye view coordinates to obtain global vehicle trajectory data and high-precision lane-level map data corresponding to the road section;

[0019] S122. According to the method of step S121, global vehicle trajectory data and high-precision lane-level map data corresponding to different road sections are collected, and all collected data are combined as a data set.

[0020] Furthermore, the specific process of step S121 is as follows:

[0021] The roadside sensing unit first collects static map information of the road section, including the drivable area, lane lines, and the bifurcation and aggregation relationship between lanes within the section area;

[0022] According to the collection frequency of the roadside perception unit, the original traffic data on the road section is collected frame by frame, and the position of all vehicles on the road is obtained through the target detection and tracking algorithm. The vehicle information passing through the field of view of the roadside perception unit is recorded, including the frame number, vehicle number, and vehicle position information;

[0023] After the recording is completed, the vehicle's location information and static map information are converted into two-dimensional bird's-eye view coordinates based on the road plane through the internal parameters of the roadside perception unit, and the global vehicle trajectory data and high-precision lane-level map data for the road section are obtained.

[0024] Furthermore, the step S13 specifically includes the following steps:

[0025] S131. Assign different scenarios to a test set and a preliminary training set according to whether the collected vehicle trajectory data is complete, wherein scenarios with missing vehicle trajectories are assigned to the test set, and the remaining scenarios are assigned to the preliminary training set;

[0026] S132, traversing each scene of the preliminary training set, masking part of the tracks of some vehicles in each scene, taking the masked data as the data input and the data before the mask as the true value, that is, defining the data before the mask and the data after the mask as the data to be repaired and the true data respectively;

[0027] Afterwards, the preliminary training set is randomly divided into two parts according to the set ratio, which belong to the final training set and the validation set respectively.

[0028] Furthermore, the vehicle trajectory defect specifically refers to the presence of vehicle trajectory interruption in the global vehicle trajectory data.

[0029] Furthermore, the vehicle trajectory missing data repair model constructed in step S2 includes a TrajGAT (Trajectory and Lane Graph Attention Network) network, wherein the TrajGAT network is based on the input lane graph data to learn the topological relationship of the lane graph data, and divides the trajectory data into two parts according to the time when the vehicle has missing data, and inputs them into two encoders respectively, so as to output the time series features of each vehicle in the two parts;

[0030] The output results of the two encoders are input into two fusion modules. The TrajGAT network is used to exchange data between the temporal features of each vehicle and the lane map. The attention mechanism is then used to exchange data between vehicles. The results of the two parts are fused based on the attention mechanism to output the final trajectory defect repair results of each vehicle.

[0031] Furthermore, the step S2 specifically includes the following steps:

[0032] S21. Establish a vehicle trajectory missing data repair model based on graph neural network and attention mechanism. The input of the model is the missing vehicle trajectory data in a scene and its corresponding high-precision lane-level map data. The specific expression formula of this data set is:

[0033] R input =[Traj; G]

[0034]

[0035] Among them, R input is a composite data set of input, which consists of two parts: Traj is the trajectory data of all vehicles, and G is the lane map data converted into a graph structure; 0 is the time when data collection starts, t 1 is the time when a vehicle has missing data, t 2 The time when data collection ends;

[0036] S22: input the data sequence obtained in step S21 into the model to repair the missing data of the global vehicle motion trajectory, and output the data from t 1 +1 time to t 2 The specific expression formula of the output of K multimodal trajectory data at time is:

[0037]

[0038] Among them, Traj output is the patched trajectory information of all vehicles. is the patch value of all vehicle positions in the scene at time t, and k is the different patch values ​​in the multimodal case.

[0039] Furthermore, the specific process of iterative training in step S3 is: randomly extracting data corresponding to the scene in the training set and inputting it into the vehicle trajectory missing data repair model for inference, calculating the loss function and performing back propagation, wherein the training loss function is the minimum difference between the repaired trajectory and the true value, so that the iterative training obtains the vehicle trajectory missing data repair model with the optimal parameters.

[0040] Compared with the prior art, the present invention proposes a vehicle trajectory missing data repair method specifically for roadside perception units. It utilizes the motion correlation of adjacent vehicles in a road section and the lane-level map data in the corresponding scene, and based on a graph neural network and an attention mechanism, constructs a vehicle trajectory missing data repair model, thereby being able to repair the vehicle missing trajectory perceived in an ordinary road environment, and achieve accurate and stable vehicle missing trajectory data repair.

[0041] The present invention collects global vehicle trajectory data and corresponding high-precision lane-level map data in different road sections, that is, in different scenarios, and preprocesses the data to obtain training set, verification set and test set data. The constructed vehicle trajectory missing data repair model uses graph neural network to learn the topological structure of map information, uses attention mechanism to transfer information between map and vehicle, and learns the mutual influence relationship between vehicles. The vehicle trajectory missing data repair model divides the input vehicle trajectory information into two parts in chronological order: from the beginning of the data until there is missing data of the vehicle, and from the occurrence of missing data of the vehicle until the end of the data. The vehicle trajectory missing data repair model uses two fusion modules to process the two parts of data respectively, and then uses the attention mechanism to aggregate the output results of the two fusion modules. Finally, the final global vehicle trajectory repair result is obtained through the linear residual layer, thereby effectively ensuring the accuracy of the repair, so that it can adapt to the global vehicle missing motion trajectory repair in different traffic scenarios.

[0042] The present invention collects data based on existing roadside perception units. It only takes a small amount of cost to collect lane-level maps of the perception range of each device, so the present invention can be quickly deployed on existing roadside perception units, saving investment in road facilities. The accurate, stable, and high-quality vehicle trajectory repair results obtained by the repair method of the present invention can more accurately perceive and judge the traffic conditions of the road, and can be used as one of the high-quality data sources for applications such as autonomous driving and vehicle-road collaboration, thereby effectively improving the operation efficiency and safety of the road. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0044] Figure 2 It is a schematic diagram of the application process of the embodiment;

[0045] Figure 3 A schematic diagram of repairing vehicle defective trajectory data in an embodiment;

[0046] Figure 4 A schematic diagram of a vehicle trajectory defect data repair model based on an attention mechanism and a lane-level map constructed by the present invention;

[0047] Figure 5 Schematic diagram of comparison between the multi-modal vehicle defect trajectory repair result and the real trajectory in the embodiment. DETAILED DESCRIPTION

[0048] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] Example

[0050] like Figure 1 As shown, a method for repairing vehicle trajectory defective data for a roadside perception unit includes the following steps:

[0051] S1. Collect global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios and merge them as a data set. By preprocessing the data in the data set, the data set is divided into a training set, a validation set, and a test set. The global vehicle trajectory data refers to the spatiotemporal motion trajectory set composed of the motion trajectory of each individual vehicle when multiple vehicles move simultaneously within a certain range. Specifically:

[0052] S11, selecting different road sections, and taking each road section as a scene;

[0053] S12, collecting global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios to combine as a data set, firstly, using a roadside sensing unit to collect trajectory data of all vehicles passing through a road section within a set time period and a lane-level map of the road section;

[0054] The collected vehicle trajectory data and lane-level map are then converted into ground-based two-dimensional bird's-eye view coordinates to obtain global vehicle trajectory data and high-precision lane-level map data corresponding to the road section;

[0055] According to the above data collection method, global vehicle trajectory data and high-precision lane-level map data corresponding to different road sections are collected, and all collected data are combined as a data set;

[0056] In the process of data collection, the static map information of the road section is first collected through the roadside perception unit, including the drivable area, lane lines, and the bifurcation and aggregation relationship between lanes in the area of ​​the road section;

[0057] According to the collection frequency of the roadside perception unit, the original traffic data on the road section is collected frame by frame, and the position of all vehicles on the road is obtained through the target detection and tracking algorithm. The vehicle information passing through the field of view of the roadside perception unit is recorded, including the frame number, vehicle number, and vehicle position information;

[0058] After recording is completed, the vehicle's location information and static map information are converted into two-dimensional bird's-eye view coordinates based on the road plane through the internal parameters of the roadside perception unit, and the global vehicle trajectory data and high-precision lane-level map data of the road section are obtained;

[0059] S13, taking the data corresponding to the scene with missing vehicle trajectories in the data set (missing vehicle trajectories refer to the situation where the trajectories of one or more vehicles are interrupted due to adverse factors such as occlusion or insufficient algorithm detection capability in the global vehicle trajectory) as the test set, and then preprocessing the data corresponding to the remaining scenes to obtain a training set and a validation set. First, according to whether the collected vehicle trajectory data is complete, different scenes are assigned to the test set and the preliminary training set, wherein the scene with missing vehicle trajectories belongs to the test set, and the remaining scenes belong to the preliminary training set;

[0060] Then, each scene of the preliminary training set is traversed, and part of the tracks of some vehicles in each scene are masked. The masked data is used as the data input, and the data before the mask is used as the true value. That is, the data before the mask and the data after the mask are defined as the data to be repaired and the true data respectively.

[0061] Then the preliminary training set is randomly divided into two parts according to the set ratio, one for the final training set and the other for the validation set;

[0062] S2. Based on graph neural network and attention mechanism, a vehicle trajectory missing data patching model is constructed. The vehicle trajectory missing data patching model includes TrajGAT network. TrajGAT network is based on the input lane map data to learn the topological relationship of lane map data, and divides the trajectory data into two parts according to the time when the vehicle has missing data. The two parts are input into two encoders respectively to output the time series features of each vehicle in the two parts.

[0063] The output results of the two encoders are input into two fusion modules. The TrajGAT network is used to exchange data between the time series features of each vehicle and the lane map. The attention mechanism is then used to exchange data between vehicles. The results of the two parts are fused based on the attention mechanism to output the final trajectory defect repair results of each vehicle.

[0064] It can be seen that step S2 specifically includes the following steps:

[0065] S21. Establish a vehicle trajectory missing data repair model based on graph neural network and attention mechanism. The input of the model is the missing vehicle trajectory data in a scene and its corresponding high-precision lane-level map data. The specific expression formula of this data set is:

[0066] R input =[Traj; G]

[0067]

[0068] Among them, R input is a composite data set of input, which consists of two parts: Traj is the trajectory data of all vehicles, and G is the lane map data converted into a graph structure; 0 is the time when data collection starts, t 1 is the time when a vehicle has missing data, t 2 The time when data collection ends;

[0069] S22: input the data sequence obtained in step S21 into the model to repair the missing data of the global vehicle motion trajectory, and output the data from t 1 +1 time to t 2 The specific expression formula of the output of K multimodal trajectory data at time is:

[0070]

[0071] Among them, Traj output is the patched trajectory information of all vehicles. is the patch value of all vehicle positions in the scene at time t, and k is the different patch values ​​in the multimodal case;

[0072] S3, iteratively train the vehicle trajectory defective data patching model using the training set, and use the verification set to test the patching effect of the vehicle trajectory defective data patching model to obtain the vehicle trajectory defective data patching model with optimal parameters. When performing iterative training, randomly extract data corresponding to the scene in the training set and input it into the vehicle trajectory defective data patching model for inference, calculate the loss function and perform back propagation, wherein the training loss function is the minimum difference between the patched trajectory and the true value, and the vehicle trajectory defective data patching model with optimal parameters is obtained by iterative training in this way;

[0073] S4. Input the test set or the vehicle trajectory data to be repaired into the vehicle trajectory defective data repair model with optimal parameters, and output the corresponding repaired global vehicle trajectory complete data.

[0074] This embodiment applies the above method, such as Figure 2 As shown, it mainly includes the following contents:

[0075] Step 1: Select a road section and use roadside high-definition cameras, millimeter-wave radars, lidar and other sensors and target detection and tracking algorithms to collect trajectory data of all vehicles passing through the section within a period of time, as well as the lane-level map of the section. The collected trajectory and map data are converted into two-dimensional ground-based bird's-eye view coordinates to obtain the global vehicle motion trajectory data and high-precision lane-level map of the section.

[0076] In this embodiment, the Argoverse, a public dataset with an existing collection method similar to step one, is used as the basic source of vehicle trajectory spatiotemporal information data. The data in Argoverse comes from a subset of data collected by Argo AI's autonomous driving test vehicles in Miami and Pittsburgh (these two American cities face different urban driving challenges and local driving habits). It includes records of sensor data across different seasons, weather conditions, and different times of the day to provide a wide range of actual driving scenarios. Argoverse is the first dataset to contain high-precision maps, which contains 290KM of high-precision map data with geometric and semantic information. The research team collected all the data using a Ford hybrid fleet fully integrated with Argo AI's autonomous driving technology. Data from two LiDARs, seven high-definition cameras, two front-facing stereo cameras, and high-precision maps were used, and all data were aligned with high-precision map data.

[0077] In order to make every scene in the global vehicle motion trajectory dataset meaningful, the present invention samples some complex and difficult traffic scenes sufficiently, monitors objects in Miami and Pittsburgh for 1006 driving hours, and mines the typical scene vehicle behavior data for up to 320 hours. These vehicles are either at intersections, turning left or right, changing lanes to adjacent lanes, or congested in heavy traffic. A total of 324,557 5-second scene sequences were collected and used as the global vehicle motion trajectory dataset. Each scene contains a 2D bird's-eye view center of each tracked object sampled at 10Hz.

[0078] Step 2: According to the method described in step 1, collect global vehicle motion trajectory data and high-precision lane-level maps of different road sections, take each road section as a scene, merge the global vehicle motion trajectory data and high-precision lane-level maps collected in all scenes as a data set, and take the scene with missing vehicle trajectories as a test set;

[0079] Part of the vehicle trajectories in the remaining scenes are masked, and the data before and after masking are defined as the data to be repaired and the real data, respectively. They are randomly divided into training sets and verification sets according to the scene numbers. Multi-scene data contains a large amount of information and can provide data support for accurate repair of vehicle trajectories.

[0080] In this example, the Argoverse dataset has collected data from multiple scenes and selected 324,557 5-second scene sequences. In step 1, corresponding high-precision map data is generated for each scene. First, the scene with missing vehicle trajectories is used as the test set, and then some vehicle trajectories in the remaining scenes are masked. The data before and after masking are defined as the data to be predicted and the real data, respectively. Figure 3 As shown in the figure, the datasets were randomly divided into training set and validation set according to the scene numbers, with a ratio of 0.84:0.16.

[0081] In order to learn the representation of vehicle trajectories and complete the task of repairing vehicle missing trajectories, it is necessary to obtain vehicle trajectories under various driving behaviors. The vehicle's running trajectory is affected by many variables: its own historical trajectory, the trajectories of surrounding vehicles, the nature of the lane it is in, and other factors will affect the vehicle's driving behavior decisions and running trajectory, and these influencing factors are innumerable. To this end, sampling the vehicle's running trajectories under different combinations of situations as much as possible will help to more accurately describe the vehicle's running trajectory and complete the task of repairing the vehicle's missing trajectory data.

[0082] Step 3: Based on the data collected and sorted in step 2, the training set data is input into the model for iterative training to obtain a vehicle defective trajectory repair model (i.e., a vehicle trajectory defective data repair model) that can adapt to different traffic scenarios, such as Figure 4 As shown, the model's repair effect is tested using validation set data. The specific process is as follows: 1) The model inputs lane map data into the TrajGAT network to learn its topological relationship. According to the time when a vehicle has missing data, the trajectory data is divided into two parts, which are input into two encoders respectively, and the time series features of each vehicle in the two parts are output; 2) The results of the two encoders are sent to two fusion modules respectively, and the TrajGAT network is used to interact with the lane map data, and then the attention mechanism is used to interact with the data between vehicles; 3) The results of the two parts are fused using the attention mechanism, and the final defective trajectory repair results of each vehicle are output;

[0083] In this technical solution, the aim is to establish a neural network model that can simultaneously input the defective trajectory information of all vehicles in a scene and the corresponding map information, and output the repaired and complete trajectory information of all vehicles;

[0084] The model uses graph neural networks to learn the topological structure of map information, uses the attention mechanism to transfer information between the map and the vehicle, and learns the mutual influence relationship between vehicles;

[0085] The model divides the input vehicle trajectory information into two parts in chronological order: from the beginning of the data until a vehicle has missing data, and from the beginning of the data until the end of the data. The model uses two fusion modules to process the two parts of data respectively, and then uses the attention mechanism to aggregate the output results of the two fusion modules, and finally obtains the final global vehicle trajectory repair result through the linear residual layer.

[0086] Step 4: Based on the patching model obtained in step 3, input the test set or other vehicle trajectory data that needs to be patched into the model to obtain the patching result of the global vehicle trajectory.

[0087] In step one, it specifically includes:

[0088] Step 1.1: Select a section of the road and deploy the vehicle trajectory collection device at a fixed point on the section. Debug the device to ensure that it can capture the complete trajectory of all vehicles from entering to leaving the section. First, collect the static map information of the scene, including the drivable area, lane lines, and bifurcation and aggregation relationships between lanes in the area.

[0089] Step 1.2: Collect the original traffic data on the road frame by frame according to the collection frequency of the collection device, and obtain the position of all vehicles on the road through the target detection and tracking algorithm, and record the information of vehicles passing through the field of view of the collection device, including frame number, vehicle number, vehicle position and other information.

[0090] Step 1.3: After recording is completed, the vehicle's location information and static map information are converted into two-dimensional bird's-eye view coordinates based on the road plane by collecting the internal parameters of the equipment, and the vehicle motion trajectory dataset and high-precision lane-level map of the road section are obtained.

[0091] Accordingly, in this embodiment, a certain road section is first given, and the database provided by Argoverse is used to obtain the positions of all vehicles in the corresponding scene frame by frame, and the lane-level map data of the corresponding area is obtained according to the boundary range of the vehicle trajectory.

[0092] Then, the spatiotemporal information tensor is generated based on the vehicle serial number, frame serial number, vehicle position and other information provided by the original data. , where N represents the total number of vehicles in the scene, L represents the frame length, and 2 represents the 2 dimensions of the location information.

[0093] Since the coordinates in the Argoverse dataset have been aligned with the high-precision map, no further conversion is required. The resulting spatiotemporal tensor is packaged with the lane map data to obtain the global vehicle trajectory dataset and high-precision lane-level map for this scenario.

[0094] In step 2, it specifically includes:

[0095] Step 2.1: Using the collection method described in step 1, collect vehicle trajectory information and corresponding high-precision lane-level maps in many traffic scenarios, filter the obtained information, and assign the scenarios to training sets and test sets based on whether the obtained vehicle trajectories are complete.

[0096] Step 2.2: Traverse each scene in the training set, mask the partial tracks of some vehicles in each scene, use the masked data as the data input, and the data before the mask as the true value to calculate the error with the output of the neural network. Then randomly divide the training set into two parts according to a certain ratio, one for the final training set and the other for the validation set.

[0097] In step three, it specifically includes:

[0098] Step 3.1: Establish a global vehicle trajectory data defect repair model based on graph neural network and attention mechanism. The input of the model is the vehicle trajectory information with defects in a scene and its corresponding high-precision lane-level map. The specific expression formula of this data set is:

[0099] R input =[Traj; G]

[0100]

[0101] Among them, R input The composite data set that represents the input consists of two parts: Traj represents the trajectory information of all vehicles, and G represents the lane map information converted into a graph structure. 0 is the time when the data starts, t 1 is the time when a vehicle has missing data, t 2 The key to converting a high-precision lane-level map into a graph structure is the generation of lane nodes and their adjacency relationships. In the map corresponding to each scene, each lane line is divided into 1-meter segments, and the endpoints of each segment are used as lane nodes. The adjacency relationship between lane nodes is derived through the relationship between lane segments.

[0102] In complex traffic environments such as intersections, the difficulty of collecting vehicle trajectories will increase for roadside trajectory collection equipment due to the large number of vehicles and complex behaviors, and the vehicle trajectory is more likely to be missing. In such a complex traffic scenario, the movement of vehicles is restricted by nearby vehicles and the position of their lanes. Therefore, there is a certain spatiotemporal correlation between the trajectories of vehicles, and there is also a correlation between the trajectory of vehicles and their lane positions. In order to characterize these two correlations and complete the repair of vehicle missing trajectory data in complex traffic environments, the model uses the attention mechanism based on the python language and the pytorch deep learning framework to transfer feature information between vehicles and between vehicles and lane nodes.

[0103] Step 3.2: Input the data sequence obtained in step 3.1 into the network to patch the vehicle trajectory data and output the trajectory data of all vehicles in the scene from t 1 +1 time to t 2 K multimodal trajectory data at the time. The specific expression formula of this output is:

[0104]

[0105] Among them, Traj output is the track information of all vehicles after patching. represents the patch value of all vehicle positions in the scene at time t, and k represents different patch values ​​in multimodal situations. When the vehicle trajectory data is missing and this missing data continues to the end of the acquisition period, the vehicle's movement has a large uncertainty. To solve this problem, the model will output the corresponding trajectories of multiple possible motion behaviors of the vehicle, such as Figure 5 shown.

[0106] Step S3.3: For the vehicle trajectory data defect repair model constructed in steps 3.1 and 3.2, randomly extract scene data from the training set and input it into the model for inference, calculate the loss function and perform back propagation, where the training loss function is the minimum difference between the repaired trajectory and the true value. In this way, iterative training is performed to obtain the repair model with optimal parameters.

[0107] In step 4, based on the patching model with optimal parameters obtained in step 3, the vehicle trajectory information to be patched and the high-precision lane-level map in the corresponding scenario are input into the network to achieve accurate patching of the global vehicle motion trajectory.

[0108] In summary, this technical solution preprocesses the global vehicle motion trajectory data of multiple scenes collected by the vehicle trajectory acquisition device and the lane-level map collected by the map acquisition device according to the data preprocessing algorithm, and obtains the vehicle trajectory and lane map data set for training the deep learning algorithm; according to the representation learning method, the preprocessed global vehicle motion trajectory data and the corresponding lane map data are input into the repair model for iterative training; according to the trained global vehicle trajectory defect repair model, the defective vehicle trajectory data in different scenes is accurately and stably repaired. This technical solution is improved from the perspective of multi-dimensional representation of spatiotemporal information and multi-source fusion of data. By adding multiple types of information such as surrounding vehicles and environmental maps to increase prior knowledge, the trajectory repair effect is improved, and the accuracy and stability of the repair are improved.

Claims

1. A method for repairing vehicle trajectory defective data for a roadside perception unit, characterized in that: The following steps are involved: S1. Collect global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios to merge them into a data set, and divide the data set into a training set, a validation set, and a test set by preprocessing the data in the data set; S2. Based on graph neural network and attention mechanism, a vehicle trajectory missing data repair model is constructed; S3, iteratively training the vehicle trajectory defective data repair model using the training set, and verifying the repair effect of the vehicle trajectory defective data repair model using the verification set, so as to obtain the vehicle trajectory defective data repair model with optimal parameters; S4, inputting the test set or the vehicle trajectory data to be repaired into the vehicle trajectory defective data repair model with optimal parameters, and outputting the corresponding repaired global vehicle trajectory complete data; The vehicle trajectory missing data repair model constructed in step S2 includes a TrajGAT network, which is based on the input lane map data to learn the topological relationship of the lane map data, and divides the trajectory data into two parts according to the time when the vehicle has missing data, and inputs them into two encoders respectively to output the time series features of each vehicle in the two parts; The output results of the two encoders are input into two fusion modules. The TrajGAT network is used to exchange data between the temporal features of each vehicle and the lane map. The attention mechanism is then used to exchange data between vehicles. The results of the two parts are fused based on the attention mechanism to output the final trajectory defect repair results of each vehicle.

2. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 1, characterized in that: The global vehicle trajectory data is specifically a set of spatiotemporal motion trajectories composed of the motion trajectory of each individual vehicle when multiple individual vehicles move simultaneously within a certain range.

3. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 2, characterized in that: The step S1 specifically includes the following steps: S11, selecting different road sections, and taking each road section as a scene; S12, collecting global vehicle trajectory data and corresponding high-precision lane-level map data in different scenarios, and merging them as a data set; S13. The data corresponding to the scene with missing vehicle trajectories in the data set is used as a test set, and the data corresponding to the remaining scenes are preprocessed to obtain a training set and a verification set.

4. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 3, characterized in that: The step S12 specifically includes the following steps: S121, collecting, through a roadside sensing unit, trajectory data of all vehicles passing through a road section within a set time period and a lane-level map of the road section; The collected vehicle trajectory data and lane-level map are converted into ground-based two-dimensional bird's-eye view coordinates to obtain global vehicle trajectory data and high-precision lane-level map data corresponding to the road section; S122. According to the method of step S121, global vehicle trajectory data and high-precision lane-level map data corresponding to different road sections are collected, and all collected data are combined as a data set.

5. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 4, characterized in that: The specific process of step S121 is as follows: The roadside sensing unit first collects static map information of the road section, including the drivable area, lane lines, and the bifurcation and aggregation relationship between lanes within the section area; According to the collection frequency of the roadside perception unit, the original traffic data on the road section is collected frame by frame, and the position of all vehicles on the road is obtained through the target detection and tracking algorithm. The vehicle information passing through the field of view of the roadside perception unit is recorded, including the frame number, vehicle number, and vehicle position information; After the recording is completed, the vehicle's location information and static map information are converted into two-dimensional bird's-eye view coordinates based on the road plane through the internal parameters of the roadside perception unit, and the global vehicle trajectory data and high-precision lane-level map data for the road section are obtained.

6. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 3, characterized in that: The step S13 specifically includes the following steps: S131. Assign different scenarios to a test set and a preliminary training set according to whether the collected vehicle trajectory data is complete, wherein scenarios with missing vehicle trajectories are assigned to the test set, and the remaining scenarios are assigned to the preliminary training set; S132, traversing each scene of the preliminary training set, masking part of the tracks of some vehicles in each scene, taking the masked data as the data input and the data before the mask as the true value, that is, defining the data before the mask and the data after the mask as the data to be repaired and the true data respectively; Afterwards, the preliminary training set is randomly divided into two parts according to the set ratio, which belong to the final training set and the validation set respectively.

7. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 6, characterized in that: The vehicle trajectory defect specifically refers to the presence of vehicle trajectory interruption in the global vehicle trajectory data.

8. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 1, characterized in that: The step S2 specifically includes the following steps: S21. Establish a vehicle trajectory missing data repair model based on graph neural network and attention mechanism. The input of the model is the missing vehicle trajectory data in a scene and its corresponding high-precision lane-level map data. The specific expression formula of this data set is: R input =[Duration;G] Among them, R input It is a composite data set of input, which consists of two parts: Traj is the trajectory data of all vehicles, G is the lane map data converted into a graph structure; t0 is the time when data collection starts, t1 is the time when a vehicle has missing data, and t2 is the time when data collection ends; S22, the data sequence obtained in step S21 is input into the model to repair the missing data of the global vehicle motion trajectory, and K multimodal trajectory data of all vehicles in the scene from time t1+1 to time t2 are output. The specific expression formula of the output is: Among them, Traj output is the patched trajectory information of all vehicles. is the patch value of all vehicle positions in the scene at time t, and k is the different patch values ​​in the multimodal case.

9. The method for repairing vehicle trajectory defective data for a roadside perception unit according to claim 1, characterized in that: The specific process of the iterative training in step S3 is: randomly extracting data corresponding to the scene in the training set and inputting it into the vehicle trajectory missing data repair model for inference, calculating the loss function and performing back propagation, wherein the training loss function is the minimum difference between the repaired trajectory and the true value, so that the iterative training can obtain the vehicle trajectory missing data repair model with optimal parameters.