Track prediction method based on self-sensing mask

Through the combination of self-perception mask generation and attention neural network, the problem that trajectory prediction methods in the existing technology is difficult to learn rich features, improve the trajectory prediction accuracy and training efficiency, and realize effective trajectory prediction in the autonomous driving system .

CN119067174AActive Publication Date: 2024-12-03ZHEJIANG UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411077367.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-07
Publication Date
2024-12-03
Estimated Expiration
2044-08-07

AI Technical Summary

Technical Problem

The existing trajectory prediction methods are difficult to learn rich feature representations due to the small data dimensions, resulting in limited accuracy improvement. It is too simple or difficult to retain information in the reconstruction task based on a random mask, which affects the pre-training effect.

Method used

Using a self-perception mask-based method, a high-quality mask is generated by constructing a Transformer deep neural network and a teacher-student model, and a pre-training of trajectory reconstruction is optimized by combining the attention neural network to optimize the feature capture and training efficiency of the model.

Benefits of technology

The accuracy and training efficiency of the trajectory prediction model are improved, and a more robust trajectory prediction model can be generated, which can effectively predict the trajectory of other traffic participants in the autonomous driving system and plan safe driving paths.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119067174A_ABST
    Figure CN119067174A_ABST
Patent Text Reader

Abstract

The invention discloses a trajectory prediction method based on a self-sensing mask. The method comprises the following steps: constructing a deep neural network based on Transform; copying a deep neural network with the same structure to form a teacher-student model; a full connection layer network is added to the top layer of the model, self-sensing mask generation training is carried out on a trajectory prediction data set, and the trained student model is subsequently used as a mask generation module to generate a high-quality mask; the method comprises the following steps: constructing an attention neural network based on a Transform coding-decoding structure; adding the trained mask generation module into the attention neural network, and carrying out trajectory reconstruction pre-training on the attention neural network on a trajectory prediction data set; after pre-training is finished, trajectory prediction training is carried out on the attention neural network on a trajectory data set; and applying the trained trajectory prediction network to an automatic driving system for predicting the trajectories of other traffic participants. The method can improve the training efficiency of the model, and enables the trajectory prediction to be more accurate.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of trajectory prediction, and in particular relates to a trajectory prediction method based on self-perception mask. Background Art

[0002] With the discovery and in-depth research of deep learning theory, many fields and tasks under autonomous driving have made rapid breakthroughs and significant improvements. Among them, trajectory prediction has become one of the most urgent prediction tasks and a popular research direction due to the high accuracy and real-time requirements of autonomous driving. The trajectory prediction task requires that the prediction system predicts the coordinates of the future trajectory based on the known trajectory sequence within a certain period of time. The current mainstream trajectory prediction solutions are mainly divided into methods based on recursive neural networks, methods based on graph neural networks, and methods based on attention mechanisms.

[0003] The most famous method based on recurrent neural networks is the work "Argoverse: 3D Tracking and Forecasting With Rich Maps" published by Chang et al. at the IEEE / CVFConference on Computer Vision and Pattern Recognition conference in 2019. The method introduced in this work is an important solution in this field. By calculating the distance between the vehicle and the centerline of the lane, a long short-term memory recurrent neural network (Long Short Term Memory, LSTM) encoder-decoder network that considers map information and surrounding environment information is proposed, and the recurrent neural network is used to fully aggregate the characteristics of the map environment and the vehicle.

[0004] The most famous method based on graph neural network is the work "LaneRCNN: Distributed Representations for Graph-Centric Motion Forecasting" published by Zeng et al. at the IEEE / RSJ International Conference on Intelligent Robots and Systems in 2021. The LaneRCNN method proposed in this work obtains the representation of the local lane map of each traffic participant, and encodes the participant's past trajectory and local map topological feature information through the traffic map, and finally uses the interaction module to complete the interaction of the local lane map. The method based on the attention mechanism is the current mainstream method. Among them, the most famous one is the work "GameFormer: Game-theoretic Modeling and Learning of Transformer-based Interactive Prediction and Planning for Autonomous Driving" published by Huang et al. at the IEEE / CVF International Conference on Computer Vision in 2023. This work proposes a Transformer-based interactive prediction and planning framework GameFormer, which uses a hierarchical decoding structure to capture the interaction between vehicles. In each decoding level, the decoder uses the prediction results of the previous level and the shared environmental features to iteratively refine the interaction process between vehicles, similar to the game between vehicles in layers.

[0005] However, due to the small dimensions of the data, the existing trajectory prediction methods are difficult to learn rich feature representations, which makes the accuracy improvement of related methods relatively limited. In recent years, some pre-training methods have emerged in the fields of natural language processing and computer vision. By using some proxy tasks for pre-training before the formal training of the network, the model can learn some common features, improve the model performance, and reduce the training time based on the pre-training parameters. Therefore, related pre-training work has also appeared in the field of trajectory prediction, such as the work "Forecast-MAE: Self-supervised Pre-training for Motion Forecasting with Masked Autoencoders" published by Cheng et al. at the IEEE / CVF International Conference on Computer Vision in 2023. This work uses random masks to hide some trajectory information, and then reconstructs the complete trajectory based on the unmasked trajectory information through the Transformer network. However, the random mask-based method will make the retained trajectory information too simple or too difficult for the reconstruction task, resulting in the network failing to learn useful features. Therefore, it is necessary to find a better masking method to help improve the effect of pre-training, thereby producing a more accurate trajectory prediction model. Summary of the invention

[0006] The present invention provides a trajectory prediction method based on self-perception mask, which can effectively help the trajectory prediction model obtain key features, improve the training efficiency of the model, and make the trajectory prediction more accurate.

[0007] A trajectory prediction method based on self-aware mask, comprising the following steps:

[0008] (1) Build a Transformer-based deep neural network;

[0009] (2) Copy a deep neural network with the same structure as step (1), and the two form a teacher-student model;

[0010] (3) A fully connected layer network is added to the top layer of the teacher-student model, and self-aware mask generation training is performed on the trajectory prediction dataset. The trained student model is subsequently used as a mask generation module to produce high-quality masks;

[0011] (4) Construct an attention neural network based on the Transformer encoding-decoding structure;

[0012] (5) Add the trained mask generation module to the attention neural network, generate high-quality masks to allow the attention neural network to reconstruct trajectories, and pre-train the attention neural network for trajectory reconstruction on the trajectory prediction dataset;

[0013] (6) After the pre-training is completed, the attention neural network is trained on the trajectory dataset to obtain a high-quality trajectory prediction network;

[0014] (7) Apply the trained trajectory prediction network in the autonomous driving system to predict the trajectories of other traffic participants.

[0015] The method proposed in the present invention first constructs a deep neural network based on Transformer, which is responsible for the generation of high-quality masks; then copies a neural network with the same structure to form a teacher-student model, and trains a deep learning network containing the teacher-student model; then the trained mask generation module generates high-quality masks, and constructs an attention neural network based on the Transformer encoding-decoding structure, which is responsible for reconstructing the trajectory features covered by the mask, and pre-training is performed on the trajectory data set of the real environment; finally, a Decoder structure for predicting the trajectory is added to the top layer of the pre-trained attention network, and trained on the training data set (trajectory prediction task) to predict the future trajectory sequence corresponding to the input historical trajectory. This method can effectively reduce the impact of random masks on the pre-training effect, and optimize the efficiency and prediction accuracy of model training by improving the ability of the network to capture features and the quality of learned features in pre-training, thereby generating a more robust trajectory prediction model, which can then be used in the autonomous driving system to plan a safe driving path by predicting the trajectories of other vehicles, pedestrians and other traffic participants.

[0016] In step (1), a Transformer-based deep neural network is constructed to encode the input trajectory sequence, output the predicted reconstruction loss of the corresponding position of the trajectory sequence, and generate a high-quality mask through the predicted reconstruction loss.

[0017] In step (2), a teacher-student model is constructed to allow the network to continuously optimize the quality of mask generation through self-perception training; the student network is responsible for adjusting its own prediction parameters from the true reconstruction loss, and then using the moving average update strategy to update the teacher network using the parameters of the student network, thereby reducing the fluctuation of each update and making the model more stable; the moving average update strategy is defined as follows:

[0018] θ t =mθ t +(1-m)θ s

[0019] Among them, θt represents the weight parameter of the teacher network, θ s Represents the weight parameter of the student network, m is responsible for controlling the size of the updated weight change, and is set to 0.9 in the present invention.

[0020] In order to alleviate the problem of too little trajectory prediction data, the dataset uses the large-scale real trajectory prediction dataset Argoverse1, which can improve the training quality of the mask generation network. The trained student network is used as a mask generation module.

[0021] The specific process of step (2) is as follows:

[0022] (2-1) Construct a deep neural network identical to step (1). The two networks form a teacher-student model. The student network is responsible for predicting the reconstruction loss and is trained by calculating the loss with the true reconstruction error. The teacher network updates itself using the parameters of the student model by moving average. The teacher network then predicts the reconstruction loss of the trajectory sequence during training and uses this loss to construct a high-quality mask.

[0023] (2-2) The trajectory sequence data of shape (K, T, 2) is input to the teacher network. The teacher network first encodes the trajectory sequence data into trajectory features of (K, T, H) through the Encoder encoding structure, and then inputs the trajectory features into the Decoder decoding structure, and outputs the prediction reconstruction loss of shape (K, T); where K represents the number of moving objects, T represents the time length, and H represents the feature dimension size;

[0024] (2-3) Mask the trajectory sequence data at a ratio of 60%. In order to facilitate network training, a simple to difficult mask sampling method is adopted. In each training round, a certain proportion of high reconstruction loss positions are dynamically selected as difficult masks. The 60% quantity requirement is filled by random masking. The proportion of difficult samples in the current round is defined as:

[0025]

[0026] Wherein, α0 represents the initial difficult sample ratio, which is set to 0 in the present invention; n represents the current training round; N represents the total training round; α N represents the number of difficult samples in the last training round, which is set to 0.5 in the present invention; the difficulty mask corresponding to each training round is the position of the trajectory sequence corresponding to the largest M before the reconstruction loss, where:

[0027] M=0.6×α n T

[0028] In addition, a random sample of 0.6×(1-α n)T trajectory sequence positions are masked, and the masked trajectory features are set to 0;

[0029] (2-4) The masked trajectory features are input to the student network, the trajectory features are reconstructed through the student Encoder structure, and then the reconstructed features are input to the student Decoder structure to output the predicted reconstruction loss corresponding to each mask position.

[0030] In step (3), self-aware mask generation training is performed on the trajectory prediction dataset Argoverse1 to help the mask generation module perceive the importance distribution of the trajectories of traffic participants in the autonomous driving environment.

[0031] The specific process of step (3) is as follows:

[0032] (3-1) A fully connected layer network is added to the top layer of the teacher-student model, which is responsible for decoding the trajectory features reconstructed by the student model into trajectory sequence coordinates. For example, if the trajectory feature shape reconstructed by the student model is (K, T, H), the Decoder structure will use this feature to predict the trajectory sequence coordinates of the shape (K, T, 2), where N represents the number of moving objects, T represents the time length, H represents the feature dimension size, and the 2 in the final output dimension represents the x-coordinate and y-coordinate.

[0033] (3-2) The decoded trajectory sequence coordinates and the original trajectory sequence coordinates are used to calculate the true reconstruction loss, and the MSE loss function is used to train the student network;

[0034] (3-3) The constructed deep learning model is trained on the GPU. Each GPU calculates the reconstruction loss of 32 trajectory sequence data at a time, and 8 GPUs are trained in parallel;

[0035] (3-4) The Adam optimization algorithm was used, and the parameters β1 and β2 were set to 0.9 and 0.999 respectively. The initial learning rate was 0.001, and the learning rate was reduced to 50% after every 20 cycles. A total of 60 cycles were trained until the model converged.

[0036] In step (4), an attention neural network based on the Transformer encoder-decoder structure is constructed in order to use the Transformer global attention mechanism to better capture the feature information in the complex environment of autonomous driving and complete the pre-training of the trajectory reconstruction task with high quality. Because, after the student model is trained in step (3), a high-quality mask can be generated to cover part of the trajectory sequence data. The Transformer attention mechanism can ignore these masked data and complete the trajectory reconstruction by fusing other unmasked data features.

[0037] The specific process of step (5) is as follows:

[0038] (5-1) Use the fully connected layer network to convert the input trajectory sequence data into an H-dimensional vector; for example, if the input trajectory sequence data is in the shape of (K, T, 2), after being processed by the fully connected layer network, it becomes a trajectory feature of (K, T, H). K represents the number of moving objects, T represents the time length, and H represents the feature dimension size;

[0039] (5-2) Input the trajectory sequence data to the mask generation module to generate a high-quality mask with a shape of (K, T), set the trajectory features at the corresponding position of the mask to 0, input it to the Encoder structure of the attention neural network for feature fusion, and obtain a trajectory feature with a shape of (K, T, H). Input the feature to the Decoder structure to decode the trajectory coordinates with a shape of (K, T, 2);

[0040] (5-3) Use the MSE function to calculate the prediction loss of the predicted trajectory coordinates and the true trajectory coordinates of the mask position, and train the attention neural network;

[0041] (5-4) The constructed attention neural network is trained on the GPU. Each GPU calculates the reconstructed coordinates of 32 trajectory sequence data at a time, and 8 GPUs are trained in parallel.

[0042] (5-5) The Adam optimization algorithm was used, and the parameters β1 and β2 were set to 0.9 and 0.999 respectively. The initial learning rate was 0.005, and the learning rate was reduced to 50% after every 10 cycles for a total of 50 cycles until the model converged.

[0043] In step (6), the historical trajectory data in the trajectory dataset Nuscenes is input into the pre-trained attention neural network, which outputs the future trajectory data through the Decoder structure. The MSE loss function is used to supervise the training of the trajectory prediction task to obtain a high-quality trajectory prediction model.

[0044] The obtained high-quality trajectory prediction model can be used in the autonomous driving system to plan a safe driving path by predicting the trajectories of other vehicles, pedestrians and other traffic participants.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] 1. This paper proposes a trajectory prediction method based on self-perception mask. Through the self-perception training of the teacher-student model, the prediction of trajectory reconstruction loss is completed, making the selection of trajectory mask more interpretable and reliable.

[0047] 2. The present invention designs a simple to difficult mask generation strategy. Using simple masks in the early stage of reconstruction network training is conducive to network convergence, and increasing the proportion of high-difficulty masks in the later stage of training is conducive to the network learning valuable and more general features.

[0048] 3. The present invention has strong applicability and can directly and effectively complete the pre-training of the attention neural network by adding a self-aware mask generation module to the attention neural network, helping the network to capture common feature information in the autonomous driving environment and improving the accuracy of the final trajectory prediction and training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a trajectory prediction method based on self-perception mask according to an embodiment of the present invention;

[0050] Figure 2 Schematic diagram of the model structure of the mask generation module of the present invention;

[0051] Figure 3 This is a schematic diagram of pre-training for trajectory prediction in the present invention. DETAILED DESCRIPTION

[0052] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0053] like Figure 1 As shown, a trajectory prediction method based on self-aware mask includes the following steps:

[0054] S01, build a deep neural network based on Transformer.

[0055] The deep neural network used refers to the design of the Transformer network, keeping parameters such as network depth and network width unchanged, and the feature encoding size is set to 128 dimensions.

[0056] S02, copies a deep neural network with the same structure, and the two constitute a teacher-student model.

[0057] The student network is responsible for learning from the true reconstruction loss and adjusting its own prediction parameters. Then, through the moving average update strategy, the parameters of the student network are used to update the teacher network, thereby reducing the fluctuations of each update and making the model training more stable.

[0058] S03, adds a fully connected layer network on top of the teacher-student model and trains it on the large trajectory prediction dataset Argoverse1.

[0059] Argoverse1 is a dataset of very large autonomous driving scenarios. Chang et al. published the paper "Argoverse: 3D Tracking and Forecasting with Rich Maps" at the Conference on Computer Vision and Pattern Recognition 2019. The dataset collected more than 30,000 scenes in Pittsburgh and Miami. Each scene is a trajectory sequence sampled at 10 Hz. Each sequence has a moving object, and the task is to predict its motion trajectory in the next 3 seconds given the x and y coordinates of the moving object's 2-second historical trajectory. The sequence is divided into training set, validation set and test set, with 205,942, 39,472 and 78,143 trajectory sequences respectively. This dataset is derived from a real driving environment, which enables the model to better learn the characteristics of the real environment.

[0060] The training process of the generated module is as follows Figure 2 As shown in the figure, the trajectory sequence data is first input into the teacher network. The teacher network encodes the trajectory sequence data into 128-dimensional trajectory features through the Encoder encoding structure, and then inputs the trajectory features into the Decoder decoding structure to output the predicted reconstruction loss.

[0061] Then, using a simple to difficult mask sampling method, different proportions of high reconstruction loss positions are selected for masking in each training round, and 60% of the mask quantity requirement is supplemented by randomly sampling masks, and the features of the mask positions are set to 0.

[0062] The masked trajectory features are then input into the student network to reconstruct the trajectory features, and the reconstructed features are input into the student Decoder structure to output the predicted reconstruction loss corresponding to each mask position.

[0063] Finally, the fully connected layer network is used to convert the reconstructed features into the final output trajectory coordinates, and the true reconstruction loss is calculated with the true trajectory coordinates. The true loss is used to calculate the MSE loss with the reconstruction loss predicted by the student network, thereby completing the training of the mask generation module.

[0064] S04, build an attention neural network based on the Transformer encoding-decoding structure. The attention neural network used refers to the design of the Transformer network, keeping parameters such as network depth and network width unchanged, and setting the feature dimension size to 128 dimensions.

[0065] S05, extracting the student network part in step S03 as a mask generation module to train the attention neural network in S04, the parameters of which are randomly initialized.

[0066] Training is performed on the Argoverse1 dataset. The training process at this stage is as follows: Figure 3 As shown in the figure, according to the input trajectory data, the mask generation module generates masks according to the strategy from simple to difficult, so as to obtain the masked trajectory features. The Encoder structure of the attention neural network fuses the unmasked trajectory features, and finally decodes the masked trajectory coordinates through the Decoder structure to complete the trajectory reconstruction task.

[0067] S06: Extract the attention neural network in step S05 and start training the model. The training data uses the Nuscense dataset, which was published by Holger et al. in the paper "nuScenes: A multimodal dataset for autonomous driving" at Conference on Computer Vision and Pattern Recognition 2020. The historical trajectory data is input to the model and the predicted future trajectory data is output.

[0068] Use the trained model to perform trajectory prediction tasks in the autonomous driving system, and plan a safe and reliable driving path by predicting the trajectories of other vehicles, pedestrians and other traffic participants.

[0069] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A trajectory prediction method based on self-aware mask, characterized in that: The following steps are involved: (1) Build a Transformer-based deep neural network; (2) Copy a deep neural network with the same structure as step (1), and the two form a teacher-student model; (3) A fully connected layer network is added to the top layer of the teacher-student model, and self-aware mask generation training is performed on the trajectory prediction dataset. The trained student model is subsequently used as a mask generation module to produce high-quality masks; (4) Construct an attention neural network based on the Transformer encoding-decoding structure; (5) Add the trained mask generation module to the attention neural network, generate high-quality masks to allow the attention neural network to reconstruct trajectories, and pre-train the attention neural network for trajectory reconstruction on the trajectory prediction dataset; (6) After the pre-training is completed, the attention neural network is trained on the trajectory dataset to obtain a high-quality trajectory prediction network; (7) Apply the trained trajectory prediction network in the autonomous driving system to predict the trajectories of other traffic participants.

2. The trajectory prediction method based on self-awareness mask according to claim 1 is characterized in that: In step (1), a Transformer-based deep neural network is constructed to encode the input trajectory sequence, output the predicted reconstruction loss of the corresponding position of the trajectory sequence, and generate a high-quality mask through the predicted reconstruction loss.

3. The trajectory prediction method based on self-awareness mask according to claim 1, characterized in that: In step (2), a teacher-student model is constructed to allow the network to continuously optimize the quality of mask generation through self-perception training; the student network is responsible for adjusting its own prediction parameters from the true reconstruction loss, and then using the moving average update strategy to update the teacher network using the parameters of the student network, thereby reducing the fluctuation of each update and making the model more stable; the moving average update strategy is defined as follows: i t =mθ t +(1-m)θ s Among them, θ t represents the weight parameter of the teacher network, θ s Represents the weight parameter of the student network, and m is responsible for controlling the size of the updated weight change.

4. The trajectory prediction method based on self-awareness mask according to claim 1, characterized in that: The specific process of step (2) is as follows: (2-1) Construct a deep neural network identical to step (1). The two networks form a teacher-student model. The student network is responsible for predicting the reconstruction loss and is trained by calculating the loss with the true reconstruction error. The teacher network updates itself using the parameters of the student model by moving average. The teacher network then predicts the reconstruction loss of the trajectory sequence during training and uses this loss to construct a high-quality mask. (2-2) The trajectory sequence data of shape (K, T, 2) is input to the teacher network. The teacher network first encodes the trajectory sequence data into trajectory features of (K, T, H) through the Encoder encoding structure, and then inputs the trajectory features into the Decoder decoding structure, and outputs the prediction reconstruction loss of shape (K, T); where K represents the number of moving objects, T represents the time length, and H represents the feature dimension size; (2-3) Mask the trajectory sequence data at a ratio of 60%. In order to facilitate network training, a simple to difficult mask sampling method is adopted. In each training round, a certain proportion of high reconstruction loss positions are dynamically selected as difficult masks. The 60% quantity requirement is filled by random masking. The proportion of difficult samples in the current round is defined as: Among them, α0 represents the initial difficult sample ratio; n represents the current training round; N represents the total training round; α N represents the number of difficult samples in the last training round; the difficulty mask corresponding to each training round is the position of the trajectory sequence corresponding to the M largest reconstruction loss, where: M=0.6×α n T In addition, a random sample of 0.6×(1-α n )T trajectory sequence positions are masked, and the masked trajectory features are set to 0; (2-4) The masked trajectory features are input to the student network, the trajectory features are reconstructed through the student Encoder structure, and then the reconstructed features are input to the student Decoder structure to output the predicted reconstruction loss corresponding to each mask position.

5. The trajectory prediction method based on self-aware mask according to claim 1, characterized in that: In step (3), self-aware mask generation training is performed on the trajectory prediction dataset Argoverse1 to help the mask generation module perceive the trajectory importance distribution of traffic participants in the autonomous driving environment.

6. The trajectory prediction method based on self-awareness mask according to claim 1, characterized in that: The specific process of step (3) is as follows: (3-1) A fully connected layer network is added to the top layer of the teacher-student model, which is responsible for decoding the trajectory features reconstructed by the student model into trajectory sequence coordinates; (3-2) The decoded trajectory sequence coordinates and the original trajectory sequence coordinates are used to calculate the true reconstruction loss, and the MSE loss function is used to train the student network; (3-3) The constructed deep learning model is trained on the GPU. Each GPU calculates the reconstruction loss of 32 trajectory sequence data at a time, and 8 GPUs are trained in parallel; (3-4) The Adam optimization algorithm was used, and the parameters β1 and β2 were set to 0.9 and 0.999 respectively. The initial learning rate was 0.001, and the learning rate was reduced to 50% after every 20 cycles. A total of 60 cycles were trained until the model converged.

7. The trajectory prediction method based on self-awareness mask according to claim 1, characterized in that: The specific process of step (5) is as follows: (5-1) Use a fully connected layer network to convert the input trajectory sequence data into an H-dimensional vector; (5-2) Input the trajectory sequence data to the mask generation module to generate a high-quality mask with a shape of (K, T), set the trajectory features at the corresponding position of the mask to 0, input it to the Encoder structure of the attention neural network for feature fusion, and obtain a trajectory feature with a shape of (K, T, H). Input the feature to the Decoder structure to decode the trajectory coordinates with a shape of (K, T, 2); (5-3) Use the MSE function to calculate the prediction loss of the predicted trajectory coordinates and the true trajectory coordinates of the mask position, and train the attention neural network; (5-4) The constructed attention neural network is trained on the GPU. Each GPU calculates the reconstructed coordinates of 32 trajectory sequence data at a time, and 8 GPUs are trained in parallel. (5-5) The Adam optimization algorithm was used, and the parameters β1 and β2 were set to 0.9 and 0.999 respectively. The initial learning rate was 0.005, and the learning rate was reduced to 50% after every 10 cycles for a total of 50 cycles until the model converged.

8. The trajectory prediction method based on self-awareness mask according to claim 1, characterized in that: In step (6), the historical trajectory data in the trajectory dataset Nuscenes is input into the pre-trained attention neural network, which outputs the future trajectory data through the Decoder structure. The MSE loss function is used to supervise the training of the trajectory prediction task to obtain a high-quality trajectory prediction model.

Citation Information

Patent Citations

  • Vehicle track prediction method based on deep learning

    CN115293237A

  • Multi-modal vehicle trajectory prediction model training method and device and trajectory prediction method

    CN115730637A

  • Neural network architecture training method for traffic participant motion prediction

    CN118036664A

  • Knowledge distillation based neural network training method, device, and storage medium

    WO2023212997A1