A vehicle trajectory prediction method based on hierarchical attention mechanism

Abstract

The application discloses a vehicle trajectory prediction method based on a layered attention mechanism. Firstly, a space-time attention module is introduced, the interaction information of surrounding vehicles is extracted through an attention mechanism, and the time dependence between historical information and future planning information is captured respectively. Then, a hard-soft attention module is constructed, which is used to realize two tasks: balancing historical information and future information, and learning the interaction information of multiple target vehicles at different geographical positions. The layered attention mechanism is used for vehicle trajectory prediction, which effectively combines future planning information and historical trajectory data to quantitatively capture the interaction information of surrounding vehicles. Compared with a traditional trajectory prediction algorithm based on historical information only, the application shows more excellent prediction performance.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of vehicle prediction, and in particular to a vehicle trajectory prediction method based on a hierarchical attention mechanism. BACKGROUND

[0002] Human drivers can often subconsciously predict the movement trends of surrounding vehicles during driving to ensure driving safety. By using this prediction information, drivers can react in advance to avoid collisions with surrounding vehicles. Similarly, autonomous vehicles also have this ability. However, when faced with complex traffic flow interactions and uncertain driver intentions, intelligent vehicles often perform poorly in terms of prediction performance.

[0003] In order to accurately predict the future motion of the target vehicle, multiple types of vehicle interaction patterns need to be considered. First, the historical interaction information with different types of vehicles, current trajectory prediction methods usually take the historical information of surrounding vehicles as input. However, the target vehicle trajectory prediction result is affected to different degrees by surrounding vehicles from different distances, and surrounding vehicles close to the target vehicle have a greater impact on the future actions of the target vehicle. Second, different important historical moments also need to be considered, because human drivers pay more attention to the historical information near the current moment. In addition, different planning results of the ego vehicle will also affect the future motion of the target vehicle, and the target vehicle can take aggressive or conservative decisions to avoid obstacles. Finally, target vehicles at different locations will assign different weights to past and future information, and for target vehicles far from the ego vehicle, the prediction result is less affected by the ego vehicle planning information.

[0004] Although past research has made some progress in extracting traffic flow historical interaction information, few studies have quantitatively described the influence of past and future interaction relationships on different target vehicle trajectory prediction. SUMMARY

[0005] The purpose of the present application is to provide a vehicle trajectory prediction method based on a hierarchical attention mechanism to overcome the problems of the prior art. The method incorporates the ego vehicle's planning information into the prediction network, quantitatively describes the influence of the interaction relationship between past and future information on different target vehicles through a hierarchical attention mechanism, and improves the accuracy of prediction.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] S1: Obtain the ego vehicle planning trajectory information, and obtain the historical motion trajectory of each target prediction vehicle corresponding to the ego vehicle and the historical motion trajectory of multiple vehicles around the target prediction vehicle in the world coordinate system, and then obtain the relative motion information of each target prediction vehicle at the current time, the corresponding ego vehicle relative planning motion information, and the relative motion information of multiple vehicles around the target prediction vehicle, so as to form the input motion vector at the current time;

[0008] S2: Input the input motion vector at the current time into the vehicle trajectory prediction network, and output the prediction motion trajectory corresponding to each target prediction vehicle.

[0009] In the S1, first, taking the position of each target prediction vehicle at the current time as the center, then processing the ego vehicle planning trajectory information, the historical motion trajectory of the current target prediction vehicle and the historical motion trajectory of multiple vehicles around the target prediction vehicle in the historical coordinate system, obtaining the ego vehicle relative planning trajectory corresponding to the current target prediction vehicle, the relative historical trajectory of the current target prediction vehicle and the relative historical trajectory of multiple vehicles around the target prediction vehicle, and finally obtaining the ego vehicle relative motion information corresponding to the current target prediction vehicle, the relative motion information of the current target prediction vehicle and the relative motion information of multiple vehicles around the target prediction vehicle at the current time after combining the speed and acceleration information of each.

[0010] In the S2, the vehicle trajectory prediction network includes a space-time attention module, a proximity matrix generation module, a hard-soft attention module and a motion behavior based trajectory decoder, the input motion vector is input as the input of the space-time attention module and the proximity matrix generation module, the space-time attention module is connected with the hard-soft attention module, the hard-soft attention module is connected with the proximity matrix generation module, and the hard-soft attention module is also connected with the motion behavior based trajectory decoder.

[0011] The space-time attention module includes a first long short-term memory unit, a second long short-term memory unit, a first time sequence attention layer, a second time sequence attention layer and a graph attention layer; the ego vehicle relative planning motion information in the input motion vector is input as the input of the first long short-term memory unit, the first long short-term memory unit is connected with the first time sequence attention layer, the relative motion information of each target prediction vehicle and the relative motion information of multiple vehicles around the target prediction vehicle in the input motion vector are input into the graph attention layer together, the relative motion information of each target prediction vehicle is also integrated into the output of the graph attention layer and input into the second long short-term memory unit, the second long short-term memory unit is connected with the second time sequence attention layer, and the outputs of the first time sequence attention layer and the second time sequence attention layer are integrated after channel dimension splicing and are taken as the output of the space-time attention module.

[0012] In the proximity matrix generation module, the proximity matrix is constructed according to the ego vehicle planning motion information in the input motion vector and the relative motion information of each target prediction vehicle and is sent to the hard-soft attention module, wherein the center is set to 1 and represents the current position of the ego vehicle, and the inverse distance of each target prediction vehicle from the ego vehicle is placed at the position of each target prediction vehicle at the current time.

[0013] The hard-soft attention module includes a channel attention layer and a target fusion module, the input of the hard-soft attention module is multiplied by the output after the channel attention layer, and the output is used as the input of the target fusion module, the output of the proximity matrix generation module is also used as the input of the target fusion module, and the output of the target fusion module is used as the output of the hard-soft attention module.

[0014] The target fusion module is a fully convolutional network symmetrically coupled with a target fusion layer, specifically including two convolutional layers, two deconvolutional layers and two target fusion layers, the input of the target fusion module is used as the input of the first convolutional layer, the first convolutional layer is connected with the first deconvolutional layer after the second convolutional layer, the output of the proximity matrix generation module is used as the input of the two target fusion layers, the output of the first convolutional layer and the output of the first deconvolutional layer are input into the first target fusion layer, the first target fusion layer is connected with the second deconvolutional layer, the output of the second deconvolutional layer and the input of the target fusion module are input into the second target fusion layer, and the output of the second target fusion layer is used as the output of the target fusion module.

[0015] In the motion behavior-based trajectory decoder, first, the target vectors corresponding to each target prediction vehicle are respectively input into two fully connected layers to obtain probability vectors corresponding to the lateral and longitudinal motion behaviors, then the probability vectors with the highest confidence in the lateral and longitudinal motion behavior probability vectors are selected respectively to generate lateral and longitudinal one-hot encodings, and finally, the target vectors corresponding to each target prediction vehicle are spliced with the lateral and longitudinal one-hot encodings to obtain the encoding vectors corresponding to each target prediction vehicle, and the encoding vectors corresponding to each target prediction vehicle are respectively input into long short-term memory units for decoding to obtain the prediction motion trajectories corresponding to each target prediction vehicle.

[0016] The application can improve the prediction accuracy by combining additional planning information and accurately capturing the interaction between different traffic participants.

[0017] The application has the following advantages:

[0018] 1、The application can realize quantitative description of the interaction information of the surrounding vehicles and target vehicles in the past and future through the cascaded attention structure, compared with the traditional method of extracting the historical interaction information of the traffic flow by using convolution and pooling, the method is more time-saving and labor-saving, and the prediction accuracy is improved to a certain extent.

[0019] 2、The application simultaneously incorporates the future trajectory information of the ego vehicle into the network, couples with the cascaded attention structure, informs the network of the motion trend of the ego vehicle, quantitatively describes the interaction between the past and future information, and alleviates the uncertainty of predicting the future driving behavior of the target vehicle. This operation is suitable for structured roads and complex urban road scenes, and by incorporating the planning information of the ego vehicle into the prediction, an interactive and more accurate prediction result is achieved. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 The network flowchart of the application.

[0021] Figure 2 The symmetric full convolution network diagram.

[0022] Figure 3 The target fusion module diagram.

[0023] Figure 4 The target fusion layer diagram.

[0024] Figure 5 The multi-modal trajectory prediction instance diagram. DETAILED DESCRIPTION

[0025] The application will be further described below in conjunction with the drawings and examples, so that those skilled in the art can have a deeper understanding of the application and can implement it. However, the following examples are only used to explain the application, and do not limit the application.

[0026] The embodiments of the application are as follows:

[0027] The application comprises the following steps:

[0028] S1: Obtain the planning trajectory information of the ego vehicle, and obtain the historical motion trajectory of all target prediction vehicles corresponding to the ego vehicle and the historical motion trajectory of multiple vehicles around the target prediction vehicles in the world coordinate system, and then obtain the relative motion information of each target prediction vehicle at the current time, the corresponding relative planning motion information of the ego vehicle, and the relative motion information of the multiple vehicles around the target prediction vehicles, so as to form the input motion vector at the current time; wherein the target prediction vehicle is a vehicle within a preset range of the ego vehicle (i.e. an autonomous vehicle), and the surrounding vehicle is a vehicle within the surrounding of the target prediction vehicle. In this embodiment, the preset range is 60.96*10.67 meters.

[0029] In S1, firstly, taking the position of each target predicted vehicle at the current moment as the center, the self-planned trajectory information, the historical motion trajectory of the current target predicted vehicle and the historical motion trajectories of multiple surrounding vehicles in the historical coordinate system are processed to obtain the relative planned trajectory of the self-planned vehicle corresponding to the current target predicted vehicle, the relative historical trajectory of the current target predicted vehicle and the relative historical trajectories of multiple surrounding vehicles. Finally, after combining their respective speed and acceleration information, the relative motion information of the self-planned vehicle corresponding to the current target predicted vehicle and the relative motion information of the current target predicted vehicle and the relative motion information of multiple surrounding vehicles at the current moment are obtained respectively.

[0030] S2: Input the input motion vector into the vehicle trajectory prediction network, and output the predicted motion trajectory corresponding to each target vehicle.

[0031] In S2, such as Figure 1 As shown, the vehicle trajectory prediction network includes a spatiotemporal attention module, a proximity matrix generation module, a hard and soft attention module, and a motion behavior-based trajectory decoder. The input motion vector serves as the input to the first long short-term memory unit, graph attention layer, and proximity matrix generation module of the spatiotemporal attention module. The spatiotemporal attention module is connected to the hard and soft attention module. The first target fusion module and the second target fusion module of the hard and soft attention module are connected to the proximity matrix generation module. The output of the hard and soft attention module is also connected to the motion behavior-based trajectory decoder.

[0032] The spatiotemporal attention module includes a first long short-term memory unit, a second long short-term memory unit, a first temporal attention layer, a second temporal attention layer, and a graph attention layer. The relative planned motion information of the vehicle in the input motion vector is used as the input of the first long short-term memory unit, which is connected to the first temporal attention layer. The relative motion information of each target predicted vehicle in the input motion vector, as well as the relative motion information of multiple surrounding vehicles, are input into the graph attention layer. The graph attention layer is used to extract the spatial interaction information of historical trajectory information. The relative motion information of each target predicted vehicle is also integrated with the output of the graph attention layer and input into the second long short-term memory unit, which is connected to the second temporal attention layer. The outputs of the first and second temporal attention layers are fused and used as the output of the spatiotemporal attention module. Specifically, the weighted historical encoding information and weighted planning encoding information corresponding to multiple target predicted vehicles are concatenated into target encoding in the channel dimension, completing the integration of historical trajectory information and planned trajectory information and serving as the input of the hardware and software attention modules.

[0033] The specific operations of the graph attention layer are as follows:

[0034] First, after applying three learnable linear transformations θ(), φ(), and ρ() to the relative motion information of the target vehicle and its surrounding vehicles, the query vector q is obtained. t Key vector K t Sum vector V t The formula is as follows:

[0035]

[0036] in, and S t These represent the relative motion information of the target vehicle and the relative motion information of its surrounding vehicles, respectively, and are represented as corresponding encoded vectors in this embodiment, W. θ , and W ρ The first, second, and third learnable matrices, respectively, satisfy W θ ∈R d W φ ∈R d and W ρ ∈R d R d Represents a real number field of dimension d.

[0037] Then, the attention weights are calculated:

[0038]

[0039] Where softmax represents the weight function, <, > represent the dot product operation, and α t This represents the correlation between multiple surrounding vehicle codes and the target predicted vehicle code, i.e., the attention weight. The surrounding vehicle codes are obtained by weighting the surrounding vehicle value vector using this weight.

[0040] Finally, gate linear units and long short-term memory units are used to aggregate the encoded information of each target predicted vehicle and surrounding vehicles. Then, the aggregate vector set H of the target vehicle trajectory at each time step is obtained. TV .

[0041] The specific operations of the temporal attention layer are as follows:

[0042] First, based on the aggregate vector set H of the target predicted vehicle trajectory at each time step output by the Long Short-Term Memory Unit,... TV Or the aggregate vector set H of the vehicle trajectory at each time step EV That is, the relative planning motion information of the vehicle is processed by the output of the first long short-term memory unit, and the attention weight corresponding to each moment is calculated, as shown in the following formula:

[0043]

[0044]

[0045] Where, β TV and β EV These represent the attention weights of the target vehicle and the driver vehicle, respectively. and Let the first and second weight matrices satisfy the following conditions: and R n Represents a real number field of dimension n. This represents the activation function.

[0046] The aggregate vector set H of the target vehicle trajectory at each time step is predicted using the attention weights at each time step. TV Or the aggregate vector set H of the vehicle trajectory at each time step EV After performing weighted summation, weighted historical coding information or weighted planning coding information is obtained.

[0047] In the proximity matrix generation module, a proximity matrix is ​​constructed based on the self-vehicle planning motion information and the relative motion information of each target predicted vehicle in the input motion vector, and sent to the first target fusion module and the second target fusion module of the hardware and software attention module. In the proximity matrix, the center is set to 1 and represents the self-vehicle position. The reciprocal of the distance between the target predicted vehicle and the self-vehicle at the current time is placed at the position of each target predicted vehicle.

[0048] The hardware and software attention modules include a channel attention layer and a target fusion module. The input of the hardware and software attention modules is multiplied by the channel attention layer and then used as the input of the target fusion module. The output of the proximity matrix generation module is also used as the input of the first target fusion module and the second target fusion module. The output of the target fusion module is used as the output of the hardware and software attention modules.

[0049] like Figure 2 As shown, the specific operation of the channel attention layer is as follows:

[0050] First, a tensor matrix centered on the autonomous vehicle is constructed based on the output of the weighted historical coding information and weighted planning coding information corresponding to multiple target predicted vehicles, denoted as the initial tensor matrix.

[0051] Next, global max pooling and global average pooling operations are performed on the initial tensor matrix to obtain the channel vectors after global max pooling and global average pooling. Then, convolution operations are performed on the channel vectors after global max pooling and global average pooling, and then they are added together to obtain the first intermediate vector. Finally, the first intermediate vector is processed by the activation function to obtain the channel attention weight vector.

[0052] Finally, the initial tensor matrix is weighted in the channel dimension by using the channel attention weight vector to obtain a channel-weighted tensor matrix.

[0053] As shown in Figure 3 The target fusion module is a full convolutional network symmetrically coupled with the target fusion layer, and specifically includes two convolutional layers, two deconvolutional layers, and two target fusion layers. The input of the target fusion module is used as the input of the first convolutional layer. The first convolutional layer is connected to the first deconvolutional layer after the second convolutional layer. The output of the first convolutional layer and the output of the first deconvolutional layer are input into the first target fusion layer. The first target fusion layer splices the output of the first convolutional layer and the output of the first deconvolutional layer in the channel dimension. The first target fusion layer is connected to the second deconvolutional layer. The output of the second deconvolutional layer and the input of the target fusion module are input into the second target fusion layer. The second target fusion layer splices the output of the second deconvolutional layer and the input of the target fusion module in the channel dimension. The output of the second target fusion layer is used as the output of the target fusion module, which is also the output of the hard-soft attention module and the input of the trajectory decoder based on the motion behavior.

[0054] As shown in Figure 4 The target fusion layer specifically operates as follows:

[0055] The target fusion layer receives the proximity matrix and the convolutional spliced tensor matrix. After the convolutional spliced tensor matrix is subjected to channel dimension maximum pooling and average pooling operations, two pooling matrices are obtained. Then, the proximity matrix is subjected to a maximum pooling operation to obtain an adjusted proximity matrix. The proximity matrix is adjusted to have the same spatial dimension size as the pooling matrices after the maximum pooling operation. The two pooling matrices and the adjusted proximity matrix are spliced in the channel dimension to obtain a feature map. Then, a spatial attention weight is obtained by using a convolution and an activation function on the feature map. The spatial attention weight and the channel-weighted tensor matrix are multiplied element by element to obtain a spatial-weighted tensor matrix.

[0056] In the trajectory decoder based on the motion behavior, the target vector corresponding to each target prediction vehicle is first input into two fully connected layers to obtain the probability vectors of the corresponding lateral and longitudinal motion behaviors. Then, the probability vectors with the highest confidence in the lateral and longitudinal motion behavior probability vectors are selected respectively to generate the lateral and longitudinal one-hot encodings. Finally, the target vector corresponding to each target prediction vehicle and its lateral and longitudinal one-hot encodings are spliced to obtain the encoding vectors corresponding to each target prediction vehicle. The encoding vectors corresponding to each target prediction vehicle are input into the long short-term memory unit decoder respectively to obtain the predicted motion trajectories corresponding to each target prediction vehicle.

[0057] The results of this example on the NGSIM dataset are shown in Figure 5

[0058] To qualitatively evaluate the accuracy of the predictions, visualizations were performed to analyze the reasonableness of the expected trajectories. In Figure 5 multimodal trajectory predictions under various traffic conditions are shown. Three cases are considered: staying in the lane, turning to the left lane, and turning to the right lane. The results show that the cascaded attention network can achieve accurate predictions by effectively identifying driving intentions and capturing the inherent multimodal behavior of target vehicles.

[0059] In summary, the present application mainly realizes a vehicle trajectory prediction method based on a hierarchical attention mechanism. This method captures the interaction information of surrounding vehicles by introducing a spatio-temporal attention module, and models the temporal dependence relationship between the continuous frame historical information and the future planning information of the ego vehicle. At the same time, the hard-soft attention module is used to balance the historical information and future information, and to learn the information of the target vehicle at different positions. Compared with traditional trajectory prediction algorithms based only on historical information, the present application effectively combines planning information and historical data using a hierarchical attention mechanism, improving prediction performance and exhibiting superior performance. The vehicle trajectory prediction method based on the hierarchical attention mechanism of this invention has broad application prospects. This method can be applied in autonomous driving systems to provide more accurate trajectory prediction for vehicles, improving the intelligent decision-making and driving safety performance of vehicles. In addition, this method can also be applied to traffic management systems to optimize traffic flow and road planning, improve traffic efficiency and reduce congestion. This invention can also be applied to traffic safety monitoring systems to help monitor and prevent traffic accidents. In summary, the vehicle trajectory prediction method based on the hierarchical attention mechanism has wide practical application potential and can promote the development and progress of the intelligent transportation field.

[0060] The above implementation examples are a preferred way of the present application, but the implementation of the present application is not limited to the above examples. Any changes, modifications, substitutions, combinations, simplifications of operation that do not violate the core spirit and principles of the present application should be considered as equivalent alternative ways and included in the protection scope of the present application. This means that when implementing the present application, corresponding adjustments can be made according to specific needs in order to better adapt to actual application scenarios.​

Claims

1. A vehicle trajectory prediction method based on a hierarchical attention mechanism, characterized in that, Includes the following steps: S1: Obtain the planned trajectory information of the vehicle, as well as the historical motion trajectories of all target predicted vehicles corresponding to the vehicle and the historical motion trajectories of multiple vehicles around each target predicted vehicle in the world coordinate system. Then, obtain the relative motion information of each target predicted vehicle at the current moment and its corresponding relative planned motion information of the vehicle and the relative motion information of multiple vehicles around the target predicted vehicle, thereby forming the input motion vector at the current moment. S2: Input the current motion vector into the vehicle trajectory prediction network and output the predicted motion trajectory corresponding to each target vehicle. In S2, the vehicle trajectory prediction network includes a spatiotemporal attention module, a proximity matrix generation module, a hard and soft attention module, and a trajectory decoder based on motion behavior. The input motion vector is used as the input to the spatiotemporal attention module and the proximity matrix generation module. The spatiotemporal attention module is connected to the hard and soft attention module, the hard and soft attention module is connected to the proximity matrix generation module, and the hard and soft attention module is also connected to the trajectory decoder based on motion behavior.

2. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 1, characterized in that, In step S1, firstly, taking the position of each target predicted vehicle at the current moment as the center, then processing the self-planned trajectory information, the historical motion trajectory of the current target predicted vehicle and the historical motion trajectories of multiple surrounding vehicles in the historical coordinate system to obtain the relative planned trajectory of the self-planned vehicle corresponding to the current target predicted vehicle, the relative historical trajectory of the current target predicted vehicle and the relative historical trajectories of multiple surrounding vehicles. Finally, after combining their respective speed and acceleration information, the relative motion information of the self-planned vehicle corresponding to the current target predicted vehicle and the relative motion information of the current target predicted vehicle and the relative motion information of multiple surrounding vehicles at the current moment are obtained respectively.

3. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 1, characterized in that, The spatiotemporal attention module includes a first long short-term memory unit, a second long short-term memory unit, a first temporal attention layer, a second temporal attention layer, and a graph attention layer. The relative planned motion information of the vehicle in the input motion vector is used as the input of the first long short-term memory unit, which is connected to the first temporal attention layer. The relative motion information of each target predicted vehicle in the input motion vector, as well as the relative motion information of multiple vehicles around it, are input into the graph attention layer. The relative motion information of each target predicted vehicle is also integrated with the output of the graph attention layer and input into the second long short-term memory unit, which is connected to the second temporal attention layer. The outputs of the first and second temporal attention layers are spliced ​​and integrated in the channel dimension as the output of the spatiotemporal attention module.

4. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 1, characterized in that, In the proximity matrix generation module, a proximity matrix is ​​constructed based on the vehicle's planned motion information and the relative motion information of each target predicted vehicle in the input motion vector, and then sent to the hardware and software attention modules. In the proximity matrix, the center is set to 1 and represents the current position of the vehicle. The reciprocal of the distance between each target predicted vehicle and the vehicle is placed at the current position of each target predicted vehicle.

5. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 1, characterized in that, The hardware and software attention module includes a channel attention layer and a target fusion module. The input of the hardware and software attention module is multiplied by the channel attention layer and then the output is used as the input of the target fusion module. The output of the proximity matrix generation module is also used as the input of the target fusion module. The output of the target fusion module is used as the output of the hardware and software attention module.

6. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 5, characterized in that, The target fusion module is a symmetrically coupled fully convolutional network with a target fusion layer. Specifically, it includes two convolutional layers, two deconvolutional layers, and two target fusion layers. The input of the target fusion module is used as the input of the first convolutional layer. The first convolutional layer is connected to the first deconvolutional layer after passing through the second convolutional layer. The output of the proximity matrix generation module is used as the input of the two target fusion layers. The outputs of the first convolutional layer and the first deconvolutional layer are input together into the first target fusion layer. The first target fusion layer is connected to the second deconvolutional layer. The output of the second deconvolutional layer and the input of the target fusion module are input together into the second target fusion layer. The output of the second target fusion layer is used as the output of the target fusion module.

7. The vehicle trajectory prediction method based on a hierarchical attention mechanism according to claim 5, characterized in that, In the motion behavior-based trajectory decoder, the target vectors corresponding to each predicted vehicle are first input into two fully connected layers to obtain the probability vectors of the corresponding lateral and longitudinal motion behaviors. Then, the probability vectors with the highest confidence among the probability vectors of the lateral and longitudinal motion behaviors are selected and lateral and longitudinal one-hot codes are generated respectively. Finally, the target vectors corresponding to each predicted vehicle are concatenated with their lateral and longitudinal one-hot codes to obtain the encoding vectors corresponding to each predicted vehicle. The encoding vectors corresponding to each predicted vehicle are then input into the Long Short-Term Memory (LSTM) unit for decoding to obtain the predicted motion trajectory corresponding to each predicted vehicle.

Images