A method for predicting the motion trajectory of clustered vehicles

Abstract

This invention relates to a method for predicting the motion trajectory of clustered vehicles, comprising: collecting spatiotemporal information data of vehicle trajectories at different time periods; dividing the spatiotemporal information data of vehicle trajectories into multiple short-time segments according to time sequence, and generating behavioral data of each vehicle within each segment to obtain vehicle temporal feature data; randomly dividing the data to obtain a training set and a validation set; constructing a vehicle trajectory prediction model based on recurrent neural networks and graph neural networks; iteratively training the model using the training set and verifying the model's prediction effect using the validation set to obtain a vehicle trajectory prediction model with optimal parameters; inputting the spatiotemporal information data of the historical trajectories of clustered vehicles in the scenario to be predicted into the model with optimal parameters, and outputting the corresponding clustered vehicle trajectory prediction result. Compared with the prior art, this invention improves upon the method from the perspective of dynamically representing vehicle interaction, enabling a more accurate description of vehicle cluster behavior and improving the accuracy and stability of clustered vehicle trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of traffic information processing technology, and in particular to a method for predicting the movement trajectory of clustered vehicles. Background Technology

[0002] In recent years, with the widespread application of positioning and video capture devices, the acquisition of spatiotemporal trajectories has become increasingly convenient. Typical spatiotemporal trajectory data mainly includes positioning data generated by vehicle-mounted or mobile devices and traffic flow data captured by roadside monitoring equipment. When multiple vehicles exist within a certain range of a road segment, their spatiotemporal trajectory data constitutes a cluster trajectory. In fields such as autonomous driving, vehicle-to-infrastructure (V2I) communication, and urban planning, this cluster trajectory data holds immense value.

[0003] In the context of busy highway traffic, vehicles are the primary moving entities, and they always appear in clusters. Each vehicle in the cluster needs to plan its own path based on information such as the position, speed, and size of other vehicles, and its behavior also affects other vehicles. For each moving entity, it is necessary to predict the movement trajectories of vehicles in the cluster over a period of time in order to maintain a safe distance and eliminate traffic safety risks. Therefore, the accuracy of cluster vehicle trajectory prediction is crucial to the decision-making system of the moving entities and is the foundation for the full-element digitalization of dynamic traffic systems.

[0004] Vehicle swarm trajectory prediction is a complex spatiotemporal sequence extrapolation task. A spatiotemporal sequence is a special type of time series that refers to the time series obtained by observing the spatial position of an object as it changes over time. Not only do individual vehicles have inherently different movement habits, but there are also interactions between vehicles in a swarm environment, and the movement patterns of individuals are implicitly influenced by the entire swarm. Currently, deep learning is a commonly used automated data analysis method for complex tasks. Learning the inherent patterns of complex interactions within swarm trajectories from the perspective of data samples using deep learning is a feasible approach. From the perspective of deep learning, the main challenges of swarm trajectory prediction include: dynamics, temporal correlation, local interactivity, and the diversity of extrapolation.

[0005] In existing technologies, research on cluster trajectory prediction includes traditional models represented by the Social Force (SF) model and spatiotemporal sequence inference models based on machine learning. Among these, machine learning-based methods such as CS-LSTM, STGCN, and DCRNN aggregate temporal and spatial information, are widely used and highly practical. However, these methods often confound ordered temporal information into a set of unordered feature vectors. Furthermore, the spatial information aggregation based on this unordered vectors fails to model the dynamics of spatial interactions, resulting in lower inference accuracy in microscopic scenarios. On the other hand, vehicle trajectory inference models, such as CS-LSTM, do not treat the movement of each vehicle in the scene equally, leading to insufficient description of local interactions. Especially in busy highway environments, all of these factors contribute to poor prediction performance and stability. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art by providing a method for predicting the trajectory of a group of vehicles, so as to accurately and stably predict the trajectory of a group of vehicles.

[0007] The objective of this invention can be achieved through the following technical solution: a method for predicting the trajectory of clustered vehicles, comprising the following steps:

[0008] S1. Collect vehicle trajectory spatiotemporal information data in different time periods, divide the vehicle trajectory spatiotemporal information data into multiple short time segments according to time sequence, and generate behavior data of each vehicle in each segment to obtain vehicle time sequence feature data, and obtain training set and validation set by random partitioning;

[0009] S2. Construct a vehicle trajectory prediction model based on recurrent neural networks and graph neural networks;

[0010] S3. Iteratively train the vehicle trajectory prediction model using the training set and verify the prediction effect of the vehicle trajectory prediction model using the validation set to obtain the vehicle trajectory prediction model with optimal parameters.

[0011] S4. Input the spatiotemporal information data of the historical trajectories of the clustered vehicles in the scenario to be predicted into the vehicle trajectory prediction model with optimal parameters, and output the corresponding clustered vehicle trajectory prediction results.

[0012] Furthermore, the vehicle trajectory spatiotemporal information data includes vehicle time information and spatial information between vehicles. Specifically, the time information is the trajectory feature information of each vehicle in the cluster running in the past time period. The trajectory feature information includes position, speed, acceleration, vehicle length, vehicle width and vehicle category information.

[0013] The spatial information between vehicles refers to the relative positions of vehicles within the vehicle cluster over a past time period.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] S11. Select a road segment and collect the feature information of all vehicles passing through the road segment frame by frame within a set time period. The feature information of the vehicles includes frame number, vehicle number, vehicle position, vehicle length, vehicle width, vehicle type, and lane number of the vehicle.

[0016] S12. Convert the vehicle position into a two-dimensional bird's-eye view coordinate with the road as the plane to obtain the spatiotemporal information data of the trajectory of all vehicles in the road segment.

[0017] S13. Divide the vehicle trajectory spatiotemporal information data into multiple short segments according to time sequence, and generate the behavior data of each vehicle in each segment to obtain vehicle temporal feature data.

[0018] S14. Randomly divide the vehicle time-series feature data to obtain a training set and a validation set.

[0019] Furthermore, the specific process of dividing the vehicle trajectory spatiotemporal information data into multiple short-time segments in step S13 according to time sequence is as follows: given a sampling length T Sample and sampling interval T Interval ;

[0020] Iterate through each frame t0 of the long sequence, assigning the number t∈[t0,t0+T] Sample ], and t mod T Interval The trajectory feature information of all vehicles in the frame with =0 is extracted to form a short time segment, and each vehicle is reassigned a sequence number according to the order of appearance.

[0021] In a short segment, the vehicles appearing in each frame may be different. Therefore, the union of the vehicles in all frames of the segment is taken, and the data part of the frames in which the vehicles do not appear is filled with 0.

[0022] Furthermore, the vehicle trajectory prediction model constructed in step S2 includes an encoding layer, a behavior inference layer, and a decoding layer. The input to the vehicle trajectory prediction model is the set of all vehicle trajectory data Traj in a certain scene. history Its spatiotemporal information is obtained through the coding layer;

[0023] The behavior inference layer uses the encoded information to predict the behavior of each vehicle;

[0024] Based on the outputs of the encoding layer and the behavior inference layer, the decoding layer predicts and outputs the trajectories Traj for all vehicles. pred .

[0025] Furthermore, both the encoding layer and the decoding layer are gated dynamic graph convolutional layers composed of dynamic graph convolutional gated neural units;

[0026] The behavior inference layer is a fully connected layer.

[0027] Furthermore, the dynamic graph convolutional gated neural unit is specifically a neural network that simultaneously uses a multi-head attention dynamic adjacency matrix, graph convolution, and recurrent neural network to learn dynamic spatiotemporal dependencies. This network inputs data into the DyGCGRU unit, uses the multi-head attention dynamic adjacency matrix to obtain the attention coefficients between vehicles, and replaces two fully connected layers with graph convolutional layers, merging convolutional operations to generate reset gates and update gates. The adjacency matrix of the convolution is the matrix composed of attention coefficients.

[0028] The multi-head attention dynamic adjacency matrix is ​​a neural network that generates attention coefficients between vehicles by using the feature information of multiple vehicles. The network embeds the input data through a fully connected layer, then uses the dot score method to score the relationship, and obtains the relationship between the vehicle and itself based on self-bias correction. Multi-head information is obtained through a multi-head fusion method.

[0029] Furthermore, the collection of vehicle trajectory data Traj history Specifically:

[0030] Traj history =[X (1) ,X (2) ,X (3) ,…,X (t) ,…,X (L) ]

[0031]

[0032] Among them, X (t) Let be the set of instantaneous trajectory data for all vehicles at time t. Let L be the instantaneous trajectory data of vehicle number n at time t, where L is the frame length of the cluster vehicle trajectory data, and N is the number of vehicles in the cluster.

[0033] Furthermore, the decoding layer predicts and outputs the trajectories Traj for all vehicles. pred Specifically:

[0034]

[0035]

[0036] in, Let be the set of instantaneous position data for all vehicles at time t. Here is the instantaneous position data of vehicle number n at time t, and H is the time length to be predicted.

[0037] Furthermore, the specific process of iterative training in step S3 is as follows: randomly sample vehicle time-series feature data from the training set and input them into the vehicle trajectory prediction model for inference, calculate the loss function and perform backpropagation, wherein the training loss function is the sum of trajectory prediction loss and behavior classification loss, and so on, iterative training yields the vehicle trajectory prediction model with optimal parameters.

[0038] Compared with existing technologies, this invention proposes a method for predicting the motion trajectory of clustered vehicles for holographic perception in smart roads. Based on machine learning methods, a vehicle cluster trajectory prediction model is designed that can simultaneously learn time series and spatial relationships. Preprocessed vehicle cluster trajectory feature data is input into the prediction model for iterative training. Then, based on the trained vehicle cluster trajectory prediction model, the vehicle cluster trajectory in different scenarios is predicted. This invention uses dynamic attention mechanism, recurrent neural network, and graph neural network methods to infer and predict the vehicle cluster trajectory in future moments using historical vehicle cluster trajectory feature data. This can more effectively utilize the dynamic interaction relationships in the vehicle cluster, thereby effectively improving the accuracy of the inference algorithm and achieving efficient, accurate, and stable vehicle cluster trajectory prediction. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0040] Figure 2 This is a schematic diagram illustrating the application process of an example.

[0041] Figure 3 A schematic diagram illustrating the establishment of a multi-head attention dynamic adjacency matrix for an example embodiment;

[0042] Figure 4 A schematic diagram illustrating the establishment of a DyGCGRU unit for an embodiment;

[0043] Figure 5 This is a schematic diagram illustrating the construction of a multi-vehicle spatiotemporal trajectory prediction model in an embodiment;

[0044] Figure 6 This is a schematic diagram comparing the vehicle trajectory prediction results with the actual speed in the example. Detailed Implementation

[0045] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments.

[0046] Example

[0047] like Figure 1 As shown, a method for predicting the trajectory of clustered vehicles includes the following steps:

[0048] S1. Collect vehicle trajectory spatiotemporal information data at different time periods (vehicle trajectory spatiotemporal information data includes vehicle time information and spatial information between vehicles; the time information specifically refers to the trajectory feature information of each vehicle in the cluster during the past time period, including position, speed, acceleration, vehicle length, vehicle width, and vehicle category information; the spatial information between vehicles refers to the relative position information between vehicles in the vehicle cluster during the past time period). Divide the vehicle trajectory spatiotemporal information data into multiple short-time segments according to time sequence, and generate the behavior data of each vehicle in each segment to obtain vehicle time-series feature data. Randomly partition the data to obtain training and validation sets. Specifically:

[0049] First, select a road segment and collect the feature information of all vehicles passing through the road segment frame by frame within a set time period. The vehicle feature information includes frame number, vehicle number, vehicle position, vehicle length, vehicle width, vehicle type, and lane number of the vehicle.

[0050] The vehicle positions are then converted into two-dimensional bird's-eye view coordinates with the road as the plane to obtain the spatiotemporal information data of the trajectories of all vehicles within the road segment.

[0051] The vehicle trajectory spatiotemporal information data is then divided into multiple short-time segments according to time sequence, and behavioral data for each vehicle within each segment is generated to obtain vehicle temporal feature data. The process of segmenting short-time segments includes:

[0052] Given sampling length T sample and sampling interval T Interval ;

[0053] Iterate through each frame t0 of the long sequence, assigning the number t∈[t0,t0+T] Sample ], and t mod T Interval The trajectory feature information of all vehicles in the frame with =0 is extracted to form a short time segment, and each vehicle is reassigned a sequence number according to the order of appearance.

[0054] In a short segment, the vehicles appearing in each frame may be different. Therefore, the union of the vehicles in all frames of the segment is taken, and the data part of the frames in which the vehicles do not appear is filled with 0.

[0055] Then, the vehicle time-series feature data is randomly divided into training and validation sets;

[0056] S2. A vehicle trajectory prediction model is constructed based on recurrent neural networks and graph neural networks. The vehicle trajectory prediction model includes an encoding layer, a behavior inference layer, and a decoding layer. Both the encoding and decoding layers use gated dynamic graph convolutional layers composed of dynamic graph convolutional gated neural units. The behavior inference layer uses a fully connected layer. The dynamic graph convolutional gated neural unit is a neural network that simultaneously uses a multi-head attention dynamic adjacency matrix, graph convolution, and recurrent neural networks to learn dynamic spatiotemporal dependencies. This network inputs data into a DyGCGRU unit, uses a multi-head attention dynamic adjacency matrix to obtain the attention coefficients between vehicles, and replaces the two fully connected layers with graph convolutional layers, merging them to generate reset and update gate convolution operations. The adjacency matrix of the convolution is the matrix composed of attention coefficients.

[0057] Multi-head attention dynamic adjacency matrix is ​​a neural network that generates attention coefficients between vehicles by using feature information of multiple vehicles. The network embeds the input data through a fully connected layer, then uses the dot score method to score the relationship, and obtains the relationship between the vehicle and itself based on self-bias correction. Multi-head information is obtained through multi-head fusion method.

[0058] The input to the vehicle trajectory prediction model is the set of all vehicle trajectory data (Traj) in a given scene. history Its spatiotemporal information is obtained through the coding layer;

[0059] The behavior inference layer uses the encoded information to predict the behavior of each vehicle;

[0060] Based on the outputs of the encoding layer and the behavior inference layer, the decoding layer predicts and outputs the trajectories Traj for all vehicles. pred ;

[0061] Among them, the collection of vehicle trajectory data Traj history Specifically:

[0062] Traj history =[X (1) ,X (2) ,X (3) ,…,X (t) ,…,X (L) ]

[0063]

[0064] In the formula, X (t) Let be the set of instantaneous trajectory data for all vehicles at time t. Let L be the instantaneous trajectory data of vehicle number n at time t, where L is the frame length of the cluster vehicle trajectory data, and N is the number of vehicles in the cluster.

[0065] The decoding layer predicts and outputs the trajectories Traj for all vehicles.pred Specifically:

[0066]

[0067]

[0068] In the formula, Let be the set of instantaneous position data for all vehicles at time t. Here is the instantaneous position data of vehicle number n at time t, and H is the time length to be predicted;

[0069] S3. Iteratively train the vehicle trajectory prediction model using the training set and verify the prediction effect of the vehicle trajectory prediction model using the validation set to obtain the vehicle trajectory prediction model with optimal parameters. The specific process of iterative training is as follows: randomly select vehicle time-series feature data from the training set and input them into the vehicle trajectory prediction model for inference, calculate the loss function and perform backpropagation. The training loss function is the sum of trajectory prediction loss and behavior classification loss. Iterative training is then performed to obtain the vehicle trajectory prediction model with optimal parameters.

[0070] S4. Input the spatiotemporal information data of the historical trajectories of the clustered vehicles in the scenario to be predicted into the vehicle trajectory prediction model with optimal parameters, and output the corresponding clustered vehicle trajectory prediction results.

[0071] This embodiment applies the above method, such as Figure 2 As shown, it mainly includes the following:

[0072] Step 1: Select the road segment to be predicted and collect trajectory data and vehicle type information of all vehicles passing through the segment within a certain period of time using roadside high-definition cameras or LiDAR sensors. Convert the collected trajectory data into ground-based two-dimensional bird's-eye view coordinates to obtain the spatiotemporal information data of vehicle trajectories for that segment during that time period.

[0073] Step one specifically includes:

[0074] Step 1.1: Select a road segment, deploy the vehicle trajectory acquisition device at a fixed point on that road segment, and debug the device to ensure that it can capture the complete trajectory of all vehicles from entering to leaving the road segment.

[0075] Step 1.2: Record vehicle information within the field of view of the traffic status collection device frame by frame according to the collection frequency of the collection device, including frame number, vehicle number, vehicle position, vehicle length and width, vehicle type, lane number, etc.

[0076] Step 1.3: After recording is completed, the vehicle's location information is converted into two-dimensional bird's-eye view coordinates with the road as the plane by using the intrinsic parameters of the acquisition device, so as to obtain the spatiotemporal characteristic data of all vehicles passing through the road section.

[0077] In this embodiment, existing publicly available datasets with similar acquisition methods and steps are used as the basic source of vehicle trajectory spatiotemporal information data. The NGSIM (Next Generation Simulation) dataset is a dataset of driving trajectories on several highways in the United States collected by the Federal Highway Administration (FHWA). It includes the driving conditions of all vehicles on roads such as US101 and I-80 over a period of time. The raw data is road video captured by a high-altitude camera over a period of time, which is then processed into trajectory point records through target detection and tracking algorithms.

[0078] This embodiment uses the I-80 dataset. On April 13, 2005, NGSIM data collectors collected detailed vehicle trajectory data of eastbound vehicles on I-80 in the San Francisco Bay Area of ​​Emeryville, California. The specific study area was approximately 500 meters (1640 ft) long and consisted of six highway lanes, including one high-occupancy vehicle (HOV) lane. An entrance ramp was also included within the study area. A total of seven simultaneous digital cameras were mounted on the roof of a 30-story building adjacent to the highway, recording vehicles passing through the study area. Vehicle trajectory data was transcribed from the video using a custom software algorithm. This vehicle trajectory dataset provides the precise location coordinates of each vehicle within the study area at a frequency of 10 Hz, thus deriving detailed lane positions and positions relative to other vehicles. A total of 45 minutes of data, consisting of three 15-minute segments, is available from the entire I-80 dataset: 4:00 PM to 4:15 PM, 5:00 PM to 5:15 PM, and 5:15 PM to 5:30 PM. These three periods represent the accumulation period of congestion, the transition period between a non-congested and congested state, and the period of complete congestion during peak hours, respectively.

[0079] Then, according to the acquisition frequency of the traffic status acquisition device, the vehicle information passing through the field of view of the acquisition device is recorded frame by frame, including frame number, vehicle number, vehicle position, vehicle length and width, vehicle type, lane number, etc., and the information is preprocessed into a form that is easy for neural network learning.

[0080] The position, speed, and other characteristic information of a vehicle at a certain time t can be represented by a vector of length F. The specific composition of the instantaneous feature vector of this vehicle and its preprocessing method are shown in the table below.

[0081] Feature Name Feature Explanation Treatment Local_X coordinate x standardization Local_Y coordinate y standardization v_length Vehicle length standardization v_Width Vehicle width standardization v_Class Vehicle type One-hot encoding v_Vel speed Normalization v_Acc acceleration Normalization flag Is it missing? -

[0082] After recording is complete, the vehicle's location information is converted into two-dimensional bird's-eye view coordinates with the road as the plane using the intrinsic parameters of the acquisition device, thus obtaining the spatiotemporal feature data of all vehicles passing through the road segment. The I-80 dataset used in this embodiment has already completed the coordinate conversion.

[0083] Step 2: According to the method described in Step 1, collect vehicle trajectory data in different time periods, divide the vehicle trajectory data in all time ranges into short segments in chronological order, treat each segment as a scene, generate behavior data for each vehicle in each segment, obtain the processed vehicle time-series feature data, and randomly divide it into training set and validation set.

[0084] In step two, the specific process of segmenting the data is as follows: given a sampling length T Sample and sampling interval T Interval Iterate through each frame t0 of the long sequence, and assign the number t∈[t0,t0+T] to each frame. Sample ], and t mod T Interval The trajectory feature information of all vehicles in frames with a value of 0 is extracted to form a short time segment, and each vehicle is reassigned a sequence number according to its appearance order. Since the vehicles appearing in different frames within a short time segment may be different, the union of vehicles in all frames of that segment is taken, and the data portion of frames where no vehicle appears is filled with 0s.

[0085] In this embodiment, within each 45-minute data set, vehicle trajectory feature data is divided into 8-second segments. The first 3 seconds of data are used as historical trajectories, and the last 5 seconds are used as the projected trajectory. Trajectories in the dataset are sampled at a rate of 10Hz. To reduce model complexity, the sampling frequency is halved, and data is extracted frame-by-frame and input into the model. A total of 31,000 8-second segments were extracted from the I-80 dataset, with 70% randomly selected as the training set, 20% as the test set, and the remainder as the validation set. Each sample contains historical trajectory feature data and the ground truth value of the projected trajectory.

[0086] Step 3: Based on the vehicle time-series feature data obtained from the preprocessing in Step 2, input the training set data into the designed prediction model based on dynamic attention for iterative training, obtain a multi-vehicle trajectory prediction model that can adapt to different traffic scenarios, and verify the prediction effect of the model. Dynamic attention is an attention method that can reflect the interaction relationship between vehicles in the scene frame by frame.

[0087] In step three, a vehicle trajectory prediction model capable of simultaneously learning time series and spatial relationships is established. The model includes an encoding layer, a behavior inference layer, and a decoding layer. The input to the model is the set of all vehicle trajectory data (Traj) in a given scene. historyThe encoding layer obtains the spatiotemporal information of each vehicle, the behavior inference layer uses the encoded information to infer the behavior of each vehicle, and the decoding layer uses the outputs of the encoding and behavior inference layers to infer the trajectory Traj of all vehicles. pred The model's encoding and decoding layers use the same architecture, both being gated dynamic graph convolutional layers composed of dynamic graph convolutional gated neural units, while the behavior inference layer uses a fully connected layer.

[0088] like Figure 4 As shown, the Dynamic Graph Convolutional Gated Neural Unit (DGRU) used in the model is a neural network that simultaneously employs multi-head attention dynamic adjacency matrices, graph convolutions, and recurrent neural networks to learn dynamic spatiotemporal dependencies, enabling the encoding and decoding of vehicle trajectory feature information. This network inputs data into a modified GRU unit, called the DyGCGRU unit. The difference between this DyGCGRU unit and a regular GRU unit is that the network uses a multi-head attention dynamic adjacency matrix to obtain attention coefficients between vehicles. It also replaces the two fully connected layers in a regular GRU with graph convolutional layers and incorporates convolutional operations for generating reset and updating gates. The adjacency matrix of the convolution is the matrix composed of the aforementioned attention coefficients.

[0089] The multi-head attention dynamic adjacency matrix used in the model (e.g.) Figure 3 The diagram shows a neural network that generates attention coefficients between vehicles using feature information from multiple vehicles. The network embeds input data through fully connected layers, then uses the dot score method for relationship scoring, followed by self-bias correction to obtain the relationship between the vehicle and itself, and finally uses a multi-head fusion method to obtain multi-head information.

[0090] The specific process of step three is as follows:

[0091] Step 3.1: Establish the above vehicle trajectory prediction model (e.g.) Figure 5 As shown), the model's input is a set of trajectory data for all vehicles in a given scene (Traj). history If this cluster trajectory contains N vehicles and has a frame length of L, the specific formula for expressing this dataset is:

[0092] Traj history =[X (1) ,X (2) ,X (3) ,…,X (L) ]

[0093]

[0094] in This represents the instantaneous trajectory data of a specific vehicle.

[0095] Step 3.2: Input the data sequence obtained in Step 3.1 into the vehicle trajectory extrapolation model for extrapolation, and output all vehicle trajectory data within the extrapolated time period. If data with a extrapolation time length of H is required, the specific formula for the output is as follows:

[0096]

[0097]

[0098] in It indicates the instantaneous position of a vehicle.

[0099] Step 3.3: For the vehicle trajectory prediction model constructed in Steps 3.1 to 3.2, randomly sample scene data from the training set and input it into the prediction model for inference. Calculate the loss function and perform backpropagation. The training loss function is the sum of the trajectory prediction loss and the behavior classification loss. Iterate in this way to obtain the prediction model with optimal parameters.

[0100] This technical solution takes into account that in high-density traffic flow, vehicle movement is constrained by nearby vehicles, thus, there is a certain spatiotemporal correlation between vehicle trajectories. If a graph structure is used to characterize the correlation between vehicle trajectories, then this graph should be sparse, directed, weighted, and dynamic: sparsity means that generally only the movements of a few vehicles that are close to each other are correlated; directionality means that the influence between two vehicles with spatiotemporally correlated trajectories is often not equal; weighting means that the strength of the influence between any two pairs of correlated vehicles is often different; and dynamism means that as the relative positions of vehicles in the cluster change, the strength and even the direction of the relationship between vehicles may change.

[0101] Therefore, using the multi-head attention dynamic adjacency matrix generation operator to obtain a relational graph with these four characteristics, and using dynamic graph convolutional gated neural units to model vehicle interactions, will help to accurately predict vehicle cluster trajectories.

[0102] Step 4: Based on the prediction model obtained in Step 3, input the historical trajectory data of the vehicle cluster to be predicted into the model to obtain the prediction result of the vehicle cluster trajectory.

[0103] To verify the effectiveness of this technical solution, this embodiment also compares its prediction performance with several traditional algorithm models, such as... Figure 6 As shown in the figure, MAN_GCRNN is the prediction model constructed by this technical solution, which is... Figure 6 It can be seen that the mean absolute error of this technical solution increases the slowest as the prediction time increases, indicating that the prediction model proposed in this technical solution is more accurate and stable.

[0104] In summary, this technical solution, based on dynamic attention mechanism, recurrent neural network, and graph neural network, uses historical vehicle cluster trajectory feature data to predict vehicle cluster trajectories in future moments. It more effectively utilizes the dynamic interaction relationships within the vehicle cluster, improves the accuracy of the prediction algorithm, and enables efficient vehicle cluster trajectory prediction.

[0105] This technical solution is applicable to existing traffic status collection equipment on the roadside of smart roads, requiring no additional hardware cost investment and saving on expensive investments in highways. Based on the accurate and stable vehicle cluster trajectory prediction results obtained from this technical solution, it can support the holographic digitization of dynamic traffic elements, and support applications such as all-time and all-space traffic flow control and vehicle-road cooperative autonomous driving, thereby effectively improving the level of refined traffic management and operational efficiency.

Claims

1. A method for predicting the trajectory of clustered vehicles, characterized in that, Includes the following steps: S1. Collect vehicle trajectory spatiotemporal information data in different time periods, divide the vehicle trajectory spatiotemporal information data into multiple short time segments according to time sequence, and generate behavior data of each vehicle in each segment to obtain vehicle time sequence feature data, and obtain training set and validation set by random partitioning; The vehicle trajectory spatiotemporal information data includes vehicle time information and spatial information between vehicles. Specifically, the time information is the trajectory feature information of each vehicle in the cluster running in the past time period. The trajectory feature information includes position, speed, acceleration, vehicle length, vehicle width and vehicle category information. The spatial information between vehicles refers to the relative position information of vehicles in the vehicle cluster over a past time period. S2. Construct a vehicle trajectory prediction model based on recurrent neural networks and graph neural networks; S3. Iteratively train the vehicle trajectory prediction model using the training set and verify the prediction effect of the vehicle trajectory prediction model using the validation set to obtain the vehicle trajectory prediction model with optimal parameters. S4. Input the spatiotemporal information data of the historical trajectory of the cluster of vehicles in the scenario to be predicted into the vehicle trajectory prediction model with the optimal parameters, and output the corresponding cluster vehicle trajectory prediction result. The vehicle trajectory prediction model constructed in step S2 includes an encoding layer, a behavior inference layer, and a decoding layer. The input of the vehicle trajectory prediction model is the set of all vehicle trajectory data in a certain scene. Its spatiotemporal information is obtained through the coding layer; The behavior inference layer uses the encoded information to predict the behavior of each vehicle; Based on the outputs of the encoding layer and the behavior inference layer, the decoding layer predicts and outputs the trajectories of all vehicles. ; Both the encoding and decoding layers employ gated dynamic graph convolutional layers composed of dynamic graph convolutional gated neural units. Specifically, the dynamic graph convolutional gated neural unit is a neural network that simultaneously uses a multi-head attention dynamic adjacency matrix, graph convolution, and a recurrent neural network to learn dynamic spatiotemporal dependencies. This network inputs data into a DyGCGRU unit, uses a multi-head attention dynamic adjacency matrix to obtain attention coefficients between vehicles, and replaces two fully connected layers with graph convolutional layers, merging convolution operations to generate reset and update gates. The adjacency matrix of the convolution is the matrix composed of attention coefficients. The multi-head attention dynamic adjacency matrix is ​​a neural network that generates attention coefficients between vehicles by using the feature information of multiple vehicles. The network embeds the input data through a fully connected layer, then uses the dot score method to score the relationship, and obtains the relationship between the vehicle and itself based on self-bias correction. Multi-head information is obtained through a multi-head fusion method. The specific process of iterative training in step S3 is as follows: randomly sample vehicle time-series feature data from the training set and input them into the vehicle trajectory prediction model for inference, calculate the loss function and perform backpropagation. The training loss function is the sum of trajectory prediction loss and behavior classification loss. In this iterative training, the vehicle trajectory prediction model with optimal parameters is obtained.

2. The method for predicting the trajectory of clustered vehicles according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Select a road segment and collect the feature information of all vehicles passing through the road segment frame by frame within a set time period. The feature information of the vehicles includes frame number, vehicle number, vehicle position, vehicle length, vehicle width, vehicle type, and lane number of the vehicle. S12. Convert the vehicle position into a two-dimensional bird's-eye view coordinate with the road as the plane to obtain the spatiotemporal information data of the trajectory of all vehicles in the road segment. S13. Divide the vehicle trajectory spatiotemporal information data into multiple short segments according to time sequence, and generate the behavior data of each vehicle in each segment to obtain vehicle temporal feature data. S14. Randomly divide the vehicle time-series feature data to obtain a training set and a validation set.

3. The method for predicting the trajectory of clustered vehicles according to claim 2, characterized in that, The specific process of dividing the vehicle trajectory spatiotemporal information data into multiple short-time segments according to time sequence in step S13 is as follows: given a sampling length and sampling interval ; Traverse each frame of the long sequence Number ,at the same time The trajectory feature information of all vehicles in the frame is extracted to form a short time segment, and each vehicle is reassigned a sequence number according to the order of appearance. In a short segment, the vehicles appearing in each frame may be different. Therefore, the union of the vehicles in all frames of the segment is taken, and the data part of the frames in which the vehicles do not appear is filled with 0.

4. The method for predicting the trajectory of clustered vehicles according to claim 1, characterized in that, The behavior inference layer is a fully connected layer.

5. The method for predicting the trajectory of clustered vehicles according to claim 1, characterized in that, The collection of vehicle trajectory data Specifically: ， ， in, for t The collection of instantaneous trajectory data for all vehicles at any given moment. For the serial number n Vehicles in t Instantaneous trajectory data at any given moment. L The frame length of the cluster vehicle trajectory data. N This refers to the number of vehicles in the cluster.

6. The method for predicting the trajectory of clustered vehicles according to claim 5, characterized in that, The decoding layer predicts and outputs the trajectories of all vehicles. Specifically: ， ， in, for t The set of instantaneous position data of all vehicles at any given moment. For the serial number n Vehicles in t Instantaneous location data at any given moment. H The length of time that needs to be predicted.

Images