Track prediction method based on continuous learning of generative adversarial network

Abstract

The invention provides a pedestrian trajectory prediction method based on a generative adversarial continuous learning algorithm, and belongs to the technical field of continuous learning. The invention provides a novel continuous learning framework for track prediction. According to the framework, firstly, a trajectory prediction basic model based on a generative adversarial architecture is constructed; secondly, an improved elastic weight consolidation (EWC) mechanism is realized, and the minimum forgetting rate is reduced to 0.56%; meanwhile, a dynamic experience playback strategy is designed, typical samples are selected through an importance sampling mechanism for playback learning, and convergence of the model can be completed within 25 rounds; and finally, dynamically balancing the loss function of each module by adopting an adaptive learning rate adjustment strategy. According to the method, the disastrous forgetting problem of a traditional trajectory prediction model in the continuous learning process is greatly relieved, the bottleneck that knowledge migration is limited in multi-scene trajectory prediction in an existing method is broken through, and the forgetting resistance and generalization ability of the model in the continuous learning environment are improved.

Description

Technical Field

[0001] The present invention relates to the fields of continuous learning and artificial intelligence, and in particular to a trajectory prediction method based on continuous learning of a generative adversarial network, which is applied to the problem of pedestrian trajectory prediction. Background Art

[0002] In the field of machine learning and artificial intelligence, continuous learning or incremental learning is an important research topic that aims to enable the model to continuously learn new tasks without forgetting the knowledge it has learned. Since neural networks are prone to "catastrophic forgetting" when learning new tasks, that is, the model loses its memory of old tasks when learning new tasks, this problem limits its application in dynamic environments. To solve this problem, researchers have proposed a variety of methods, such as elastic weight consolidation (EWC) and experience replay strategies, which aim to maintain the stability of the model for learned tasks and promote the transfer and reuse of cross-task knowledge. .

[0003] In recent years, generative adversarial networks (GANs) have been widely used in various generative tasks, including image generation and speech synthesis. GANs use adversarial learning mechanisms to achieve the game process between the generative model and the discriminative model, so that the generative model can be continuously optimized when generating samples. The core idea is to improve the quality and diversity of generated samples through the game between the generator (G) and the discriminator (D). Introducing GANs into continuous learning is expected to strengthen the predictive ability of the generative model through adversarial training, while regularizing the learning of new tasks, thereby reducing catastrophic forgetting.

[0004] The Elastic Weight Consolidation (EWC) mechanism introduces the Fisher information matrix to evaluate the importance of model parameters, so that additional constraints are imposed on important parameters when updating the model to prevent them from changing significantly during the learning of new tasks. This approach effectively preserves the model's performance on old tasks.

[0005] Experience replay technology maintains an experience pool to store samples of historical tasks, and randomly extracts them for replay during training, thereby maintaining and transferring knowledge of old tasks. This method can improve the model's adaptability to new and old tasks, but the dynamic characteristics of the tasks and the sample selection strategy still need to be further optimized.

[0006] The combination of generative adversarial networks with elastic weight consolidation and experience replay mechanisms aims to provide a more efficient trajectory prediction model that can effectively deal with catastrophic forgetting problems when facing multiple continuous learning tasks and achieve continuous learning of new tasks. The combination of generative adversarial training, elastic weight consolidation and experience replay strategies is expected to improve the generalization ability and cross-task transfer performance of the model, thereby achieving better learning and prediction effects in practical applications. Summary of the invention

[0007] In order to overcome the problems existing in the related art to at least a certain extent, this application provides a generative adversarial continuous learning algorithm based on elastic weight consolidation and experience replay to solve the catastrophic forgetting problem in pedestrian trajectory prediction. This method utilizes the generation and discrimination mechanism of the generative adversarial network (GAN), the parameter importance constraint of elastic weight consolidation, and the knowledge reuse strategy of experience replay, combined with the needs of continuous learning, to achieve efficient incremental learning and inter-task knowledge transfer, solve the problem of traditional neural networks forgetting learned knowledge in continuous task learning, and thus improve the generalization ability and stability of the model in multiple tasks.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] Step S1: Initialize the continuous learning environment and task sequence, and establish a neural network architecture that supports incremental learning;

[0010] Step S2: construct a trajectory prediction basic model based on the generative adversarial network to extract and model the dynamic features of the trajectory;

[0011] Step S3: Implement a resilient weight consolidation mechanism to prevent catastrophic forgetting and maintain the model’s performance on learned tasks;

[0012] Step S4: Execute experience replay to promote knowledge transfer and realize knowledge reuse across task scenarios;

[0013] Step S5: Repeat steps S2 to S4 until the incremental learning process of all continuous tasks is completed.

[0014] Further, step S1 includes:

[0015] Step S1.1: Construct a continuous task sequence T = {T1, T2, ..., T n}, each task represents a trajectory prediction problem in a specific scenario, task dataset Contains historical trajectory sequences and corresponding future trajectory labels;

[0016] Step S1.2: Initialize the experience playback buffer with a capacity of M, adopt a ring queue structure, and support a dynamic update mechanism based on importance;

[0017] Step S1.3: Create a task feature dictionary to store the statistical features of each task, including the feature mean vector μ i and the covariance matrix Σ i , used to calculate the similarity between tasks;

[0018] Step S1.4: Initialize the Fisher information matrix F and parameter importance matrix Ω to prepare for the subsequent elastic weight consolidation mechanism.

[0019] Furthermore, the construction of the generative adversarial network in step S2 includes the following steps:

[0020] Step S2.1: Design a trajectory encoder, including a spatiotemporal feature encoder and a dynamic feature encoder, and implement feature fusion through the attention mechanism: h fused =Attention(h pos ,h vel ), where h pos is the position feature, h vel is the speed characteristic;

[0021] Step S2.2: Construct a Transformer-based trajectory generator G, with the input including the encoded historical trajectory features h ctx and random noise z, using autoregressive method to generate prediction trajectory;

[0022] Step S2.3: Design a temporal discriminator network D to evaluate the authenticity of the generated trajectory, including a trajectory encoding layer and a temporal consistency check module;

[0023] Step S2.4: Optimize the overall objective function L total =L adv +α·L traj +β·L motion +γ·L social , where each term represents the adversarial loss, trajectory reconstruction loss, motion constraint loss, and social interaction loss, respectively.

[0024] Furthermore, the elastic weight consolidation mechanism in step S3 includes the following steps:

[0025] Step S3.1: Calculate the Fisher information matrix Reflects the importance of model parameter θ to the current task performance;

[0026] Step S3.2: Implement an online update mechanism for parameter importance: in is the parameter variance, ∈ is the smoothing factor to prevent division by zero;

[0027] Step S3.3: Construct elastic weight constraints Prevent important parameters from changing dramatically in subsequent learning, where θ * is the parameter value of the historical task, λ EWC is the regularization hyperparameter;

[0028] Step S3.4: Setting the Hierarchical Knowledge Protection Threshold Dynamically adjust λ when the accuracy of historical tasks drops below a threshold EWC ;

[0029] Step S3.5: Implement a weight adjustment mechanism based on task similarity: Where S ij is the similarity between tasks i and j;

[0030] Step S3.6: Maintain a periodic checkpoint mechanism to support selective parameter rollback when severe catastrophic forgetting is detected: θ new =αθ current +(1-α)θ checkpoint

[0031] Furthermore, the experience replay mechanism in step S4 includes the following steps:

[0032] Step S4.1: Evaluate the sample importance score by calculating the sample gradient norm: in, represents the gradient operator for the model parameter θ, L(x i ,y i ) represents the loss function of the i-th sample, where x i is the input feature, y i is the corresponding label. Sample importance score s i Reflects the contribution of samples to the update of model parameters;

[0033] Step S4.2: Design importance-based sample storage probability: where p t is the basic storage probability positively correlated with the importance of the sample, M max is the maximum capacity of the buffer, N t Indicates the total number of samples at the current moment;

[0034] Step S4.3: Implement dynamic update of sample importance based on time decay: Where λ is the time decay factor, represents the importance score of the updated i-th sample, s i represents the historical importance score of the i-th sample, Indicates the importance score currently calculated for the i-th sample;

[0035] Step S4.4: Construct the sample redundancy calculation method for kernel density estimation: r i =∑ j≠i K(xi ,x j ), to avoid storing samples that are too similar, where K(x i ,x j ) represents the kernel function, which is used to measure the sample x i With x j The similarity between

[0036] Step S4.5: Implement a sample selection strategy based on task similarity, giving priority to retaining historical samples with high similarity to the current task;

[0037] Step S4.6: Distill the knowledge through the loss L distill =KL(p old (y|x)∥p new (y|x)) maintains historical knowledge, where p old (y|x) represents the predicted probability distribution of the old model for the output y under the input x, p new (y|x) represents the predicted probability distribution of the new model for the output y under the input x, KL(·∥·) represents the KL divergence, an asymmetric measure used to measure the difference between two probability distributions;

[0038] Step S4.7: Construct the overall loss function: L total =L current +α·L EWC +β·L replay +γ·L distill

[0039] Furthermore, the iterative optimization process in step S5 includes the following steps:

[0040] Step S5.1: Evaluate the average accuracy of the current model on all learned tasks and forgetfulness in Represents the accuracy of the model for the i-th task at time t;

[0041] Step S5.2: Dynamically adjust the elastic weight consolidation strength λ according to the evaluation results EWC and experience replay strategies;

[0042] Step S5.3: Determine whether the termination condition is met. If not, return to step S2 to continue optimization.

[0043] The technical solution provided by the embodiment of the present invention has the following beneficial effects: The present invention significantly improves the anti-forgetting ability of the model in the ETH pedestrian trajectory prediction task by introducing the elastic weight consolidation (EWC) mechanism, experience replay and generative adversarial network in the continuous learning framework. By dynamically adjusting the elastic weight consolidation strength and the experience replay strategy, catastrophic forgetting can be effectively prevented, the knowledge of historical tasks can be maintained, and the learning ability of the model can be guaranteed when new tasks arrive. Generative adversarial networks play an important role in trajectory prediction. The generalization ability of the model is improved through the adversarial training mechanism, making the model more stable in a multi-task learning environment. Finally, combining these technologies enables the pedestrian trajectory prediction task to be carried out efficiently in a complex and dynamic environment, achieving better prediction accuracy and learning efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 Schematic diagram of the flow of a trajectory prediction method based on continuous learning of a generative adversarial network.

[0045] Figure 2 This is the architecture diagram of the trajectory prediction method based on the ETH dataset.

[0046] Figure 3 This is a rendering of a trajectory prediction method based on continuous learning of a generative adversarial network. DETAILED DESCRIPTION

[0047] In order to make the technical solutions in the present invention better understood by the persons skilled in the art, the technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments.

[0048] Embodiment 1:

[0049] The trajectory prediction system based on generative adversarial continuous learning disclosed in this embodiment uses the public dataset ETH. The specific process diagram is as follows: Figure 1 This figure specifically shows the implementation process of knowledge transfer and anti-forgetting mechanism in the continuous learning process.

[0050] The ETH dataset was collected in real scenes in Zurich, Switzerland, and contains two main scenes: ETH and Hotel. This implementation uses the ETH scene, which was collected at an entrance on the campus of the Swiss Federal Institute of Technology (ETH) in Zurich, and records the movement trajectories of pedestrians in the area. The trajectory data is provided in the form of a time series in the format of [frame_number, pedestrian_ID, pos_x, pos_z, pos_y, v_x, v_z, v_y], where pos_z and v_z are not used. The world coordinate system (in meters) is used, and the homography matrix is ​​provided to convert the world coordinates to image pixel coordinates. In addition, the dataset also includes recorded group walking information, assumed destinations, and obstacle information. These characteristics of the ETH dataset provide a reliable data foundation for the trajectory prediction system and support the realization of knowledge transfer and anti-forgetting mechanisms in the process of generative adversarial continuous learning.

[0051] The trajectory prediction task based on generative adversarial continuous learning disclosed in this embodiment and the trajectory prediction system architecture based on ETH data are shown in the figure below: Figure 2 As shown, the following steps are included:

[0052] Step S1: Determine the continuous learning task sequence and initialize the environment;

[0053] Wherein, step S1 specifically includes:

[0054] Step S1.1: Construct a continuous task sequence T = {T1, T2, ..., T5}}, set the number of tasks to 5 based on the ETH dataset, and each task contains 8000 trajectory sequences, of which the training set accounts for 80% and the validation set accounts for 20%;

[0055] Step S1.2: For each task T i , whose dataset The input sequence Contains the coordinates of historical trajectory points of 8 time steps, output sequence Contains coordinates of future trajectory points for 12 time steps;

[0056] Step S1.3: Initialize the experience playback buffer M = 1000, use the ring queue data structure, and assign an importance score s to each sample i , the initial value is set to 0;

[0057] Step S1.4: Initialize the Fisher information matrix F to be a zero matrix, the parameter importance matrix Ω to be a unit matrix, and set the elastic weight coefficient λ EWC The initial value of is 0.4.

[0058] Step S2: construct a trajectory prediction model based on a generative adversarial network;

[0059] The method for constructing a generative adversarial network in step S2 specifically includes:

[0060] Step S2.1.1: Design the spatiotemporal feature extraction module of the trajectory encoder, which consists of 3 fully connected layers with dimensions of [2, 64, 128, 256], each followed by a ReLU activation function and a BatchNorm layer;

[0061] Step S2.1.2: Construct a dynamic feature encoder using a two-layer LSTM network with a hidden layer dimension of 256 and inputs the velocity and acceleration vectors of adjacent trajectory points;

[0062] Step S2.1.3: Implement an 8-head attention mechanism, each with a dimension of 64, to fuse the position feature h pos and speed characteristic h vel : h fused =MultiHeadAttention(Q=h pos ,K=h vel ,V=h vel ) Among them, Q, K, and V represent query, key, and value respectively.

[0063] Step S2.2.1: Construct a Transformer generator with 6 encoder blocks, each containing multi-head self-attention (head=8, d k =64) and a two-layer feedforward network (dim=[256,512,256]);

[0064] Step S2.2.2: Add residual connections and normalization operations LayerNorm to each encoder block of the generator to prevent gradient disappearance: h out =LayerNorm(h in +SubLayer(h in )) Among them, h in 、h out They represent the input and output features of the encoder block respectively, and SubLayer() represents the transformation function of the encoder sublayer;

[0065] Step S2.2.3: Design the decoder structure, use autoregression to generate trajectories, and use beamsearch (beam_size=5) to select the optimal path for each prediction;

[0066] Step S2.3.1: Implement the temporal discriminator and use 3 layers of 1D convolutional layers to extract features. The convolution kernel size is [3, 3, 3] and the number of channels is [64, 128, 256];

[0067] Step S2.3.2: Add a temporal attention module to the discriminator to calculate the dependencies between different time steps;

[0068] Step S2.4: Construct loss functions (adversarial loss, trajectory reconstruction loss, motion constraint loss, and social interaction loss): L total =L adv +α·L traj +β·L motion +γ·L social Among them, α=1.0, β=0.5, γ=0.3, and the losses are defined as follows: L social =max(0,d safe -||x i -x j ||2) Among them, x pred represents the trajectory position predicted by the model, x real is the true trajectory position; v t represents the speed at time t, a t represents the acceleration at time t; d safe represents the safety distance threshold between pedestrians, x i 、x j represents the positions of the i-th and j-th pedestrians.

[0069] Step S3: Implementing elastic weight consolidation mechanism;

[0070] Wherein, step S3 specifically includes:

[0071] Step S3.1.1: After each task is trained, use the validation set to calculate the Fisher information matrix: Among them, N is the number of samples in the validation set, x i is the input feature of the i-th sample, y i is the true label of the i-th sample, p(y i |x i ,θ) represents the given input x i When the model has the label y i The predicted probability of

[0072] Step S3.1.2: Set the batch_size of the Fisher information matrix to 64 and use 100 batches of data for estimation;

[0073] Step S3.2: Implement online parameter importance update: Among them, ∈ = 1e-6, Estimate parameter variances by moving average;

[0074] Step S3.3: Construct elastic weight constraints: Among them, λ EWC Dynamic adjustment based on task similarity:

[0075] Step S4: Execute the experience replay mechanism;

[0076] Wherein, step S4 specifically includes:

[0077] Step S4.1: Calculate sample importance score: Use the gradient accumulation value of the Adam optimizer for approximation;

[0078] Step S4.2: Design sample storage probability: Among them, M max The upper limit of the buffer size;

[0079] Step S4.3: Implementing a time-decay update mechanism for sample importance;

[0080] Specifically, the importance score s of a sample i It will be updated attenuated over time to ensure that recent data has a greater impact on model training. The update formula is: Among them, α is the attenuation coefficient, and its value range is [0,1]. is the importance score of the current task. In this way, the older data gradually reduces its influence on the model, making the model better adapted to the learning of new tasks;

[0081] Step S4.4: Implement knowledge distillation loss;

[0082] In order to prevent the old tasks from being forgotten, knowledge distillation is used to keep the knowledge of the old tasks. Its loss function is defined as: L distill =T 2 ·KL(σ(z old / T)||σ(z new / T)) Among them, T is the temperature hyperparameter, σ is the softmax function, and z old and z new are the outputs of the old task and the new task respectively. By minimizing the KL divergence, the model can retain the knowledge of the old task as much as possible when learning the new task, reducing the forgetting phenomenon.

[0083] Step S5: iterative optimization;

[0084] Wherein, step S5 specifically includes:

[0085] Step S5.1: Calculate the average accuracy: Among them, A t is the average accuracy of the current task, is the prediction accuracy of task i at time t;

[0086] Step S5.2: Evaluate the forgetting degree:

[0087] Step S5.3: Implement cosine annealing learning rate adjustment: Among them, η0 = 0.001 is the initial learning rate, and T = 10000 is the total number of iterations. This strategy helps to smoothly reduce the learning rate during training and improve training stability;

[0088] Step S5.4: Set the termination condition: Average accuracy A t Reaching the set threshold (such as 0.8), the forgetting degree F t When the value is less than the set threshold (such as 0.2), the maximum number of iterations is reached.

[0089] It can be seen from the above technical solutions that the present invention has the following advantages: through the carefully designed elastic weight consolidation and experience replay mechanism, the catastrophic forgetting problem in continuous learning is effectively solved. Combining the trajectory prediction model of the generative adversarial network not only improves the prediction accuracy, but also realizes the knowledge transfer across scenarios. The experimental results are as follows Figure 3 As shown in the figure, this method performs well in three key aspects: training efficiency, memory retention and prediction performance. On the ETH dataset, the accuracy can reach 90.01%, and the forgetting rate is as low as 0.56%, which can achieve fast, stable and high-precision continuous trajectory prediction.

[0090] For those skilled in the art, the present invention is not limited to the exemplary details of the above embodiments. The protection scope of the present invention should be limited by the scope of the appended claims rather than the above embodiments. Any equivalent changes made according to the technical solution of the present invention should fall within the protection scope of the present invention.

Claims

1. A trajectory prediction method based on continuous learning of a generative adversarial network, for achieving efficient continuous trajectory learning and prediction, characterized in that: The following steps are involved: Step S1: Initialize the continuous learning environment and task sequence, and establish a neural network architecture that supports incremental learning; Step S2: construct a trajectory prediction basic model based on the generative adversarial network to extract and model the dynamic features of the trajectory; Step S3: Implement a resilient weight consolidation mechanism to prevent catastrophic forgetting and maintain the model’s performance on learned tasks; Step S4: Execute experience replay to promote knowledge transfer and realize knowledge reuse across task scenarios; Step S5: Repeat steps S2 to S4 until all continuous task learning processes are completed.

2. The method according to claim 1, characterized in that: The continuous learning environment initialization in step S1 includes the following steps: Step S1.1: Construct a continuous task sequence T = {T1, T2, ..., T n }, each task represents a trajectory prediction problem in a specific scenario, task dataset Contains historical trajectory sequences and corresponding future trajectory labels; Step S1.2: Initialize the experience playback buffer with a capacity of M, adopt a ring queue structure, and support a dynamic update mechanism based on importance; Step S1.3: Create a task feature dictionary to store the statistical features of each task, including the feature mean vector μ i and the covariance matrix Σ i , used to calculate the similarity between tasks; Step S1.4: Initialize the Fisher information matrix F and parameter importance matrix Ω to prepare for the subsequent elastic weight consolidation mechanism.

3. The method according to claim 1, characterized in that: The construction of the generative adversarial network in step S2 includes the following steps: Step S2.1: Design a trajectory encoder, including a spatiotemporal feature encoder and a dynamic feature encoder, and implement feature fusion through the attention mechanism: h fused =Attention(h pos ,h vel ), where h pos is the position feature, h vel is the speed characteristic; Step S2.2: Construct a Transformer-based trajectory generator G, with the input including the encoded historical trajectory features h ctx and random noise z, using autoregressive method to generate prediction trajectory; Step S2.3: Design a temporal discriminator network D to evaluate the authenticity of the generated trajectory, including a trajectory encoding layer and a temporal consistency check module; Step S2.4: Optimize the overall objective function L total =L adv +α·L traj +β·L motion +γ·L social , where each term represents the adversarial loss, trajectory reconstruction loss, motion constraint loss, and social interaction loss, respectively.

4. The method according to claim 1, characterized in that: The elastic weight consolidation mechanism in step S3 includes the following steps: Step S3.1: Calculate the Fisher information matrix Reflects the importance of the model parameter θ to the current task performance; Step S3.2: Implement an online update mechanism for parameter importance: in is the parameter variance, ∈ is the smoothing factor to prevent division by zero; Step S3.3: Construct elastic weight constraints Prevent important parameters from changing dramatically in subsequent learning, where θ * is the parameter value of the historical task, λ EWC is the regularization hyperparameter; Step S3.4: Setting the Hierarchical Knowledge Protection Threshold Dynamically adjust λ when the accuracy of historical tasks drops below a threshold EWC ; Step S3.5: Implement a weight adjustment mechanism based on task similarity: Where S ij is the similarity between tasks i and j; Step S3.6: Maintain a periodic checkpoint mechanism to support selective parameter rollback when severe catastrophic forgetting is detected: θ new =αθ current +(1-α)θ checkpoint .

5. The method according to claim 1, characterized in that: The experience replay mechanism in step S4 includes the following steps: Step S4.1: Evaluate the sample importance score by calculating the sample gradient norm: in, represents the gradient operator for the model parameter θ, L(x i ,y i ) represents the loss function of the i-th sample, where x i is the input feature, y i is the corresponding label. Sample importance score s i Reflects the contribution of samples to the update of model parameters; Step S4.2: Design importance-based sample storage probability: where p t is the basic storage probability positively correlated with the importance of the sample, M max is the maximum capacity of the buffer, N t Indicates the total number of samples at the current moment; Step S4.3: Implement dynamic update of sample importance based on time decay: Where λ is the time decay factor, represents the importance score of the updated i-th sample, s i represents the historical importance score of the i-th sample, Represents the importance score currently calculated for the i-th sample; Step S4.4: Construct the sample redundancy calculation method for kernel density estimation: r i =∑ j≠i K(x i ,x j ), to avoid storing samples that are too similar, where K(x i ,x j ) represents the kernel function, which is used to measure the sample x i With x j The similarity between Step S4.5: Implement a sample selection strategy based on task similarity, giving priority to retaining historical samples with high similarity to the current task; Step S4.6: Distill the knowledge through the loss L distill =KL(p old (y|x)∥p new (y|x)) maintains historical knowledge, where p old (y|x) represents the predicted probability distribution of the old model for the output y under the input x, p new (y|x) represents the predicted probability distribution of the new model for the output y under the input x, KL(·∥·) represents the KL divergence, an asymmetric measure used to measure the difference between two probability distributions; Step S4.7: Construct the overall loss function: L total =L current +α·L EWC +β·L replay +γ·L distill .

6. The method according to claim 1, characterized in that: The iterative optimization process in step S5 includes the following steps: Step S5.1: Evaluate the average accuracy of the current model on all learned tasks and forgetfulness in Represents the accuracy of the model for the i-th task at time t; Step S5.2: Dynamically adjust the elastic weight consolidation strength λ according to the evaluation results EWC and experience replay strategies; Step S5.3: Determine whether the termination condition is met. If not, return to step S2 to continue optimization.

Images