Abstract:Learning to predict the trajectories of pedestrians is essential for improving safety and efficiency of mobile robots. The prediction is challenging since the robot needs to operate in multiple environments in which the motion patterns of pedestrians are different between environments. Existing pedestrian trajectory prediction models heavily rely on the availability of representative data samples during training. In the presence of additional training data from a new environment, these models must be retrained on all datasets to avoid catastrophic forgetting of the knowledge obtained from the already supported environments. In this paper, we address this catastrophic forgetting problem in the context of learning to predict the trajectories of pedestrians. We propose a pseudo-rehearsal approach based on a novel Generative Replay (GR) model, referred to as Social-GR. The proposed method is consistent with crowd motion patterns and is free of any explicit reference to past experiences. To demonstrate the problem of catastrophic forgetting and evaluate our solution, we develop the Continual Trajectory Prediction Benchmark, which consists of four tasks, each representing a real-world pedestrian trajectory dataset from a different environment. By conducting several experiments, we show that our proposed Social-GR approach significantly outperforms other continual learning methods that depend on explicit experience replay, including the state-of-the-art conditional-GR model. We further illustrate the robustness of our proposed approach to mitigating catastrophic forgetting by switching the order of environments and employing a more complex prediction model.

Continual Pedestrian Trajectory Learning With Social Generative Replay

Self-paced Learning for Pedestrian Trajectory Prediction.

Lifelong Vehicle Trajectory Prediction Framework Based on Generative Replay

Adaptive Trajectory Prediction Without Catastrophic Forgetting

Continual learning-based trajectory prediction with memory augmented networks

t-DGR: A Trajectory-Based Deep Generative Replay Method for Continual Learning in Decision Making

Improving Pedestrian Prediction Models With Self-Supervised Continual Learning

CR-LSTM: Collision-prior Guided Social Refinement for Pedestrian Trajectory Prediction

Continual Interactive Behavior Learning With Traffic Divergence Measurement: A Dynamic Gradient Scenario Memory Approach

Pedestrian Trajectory Prediction with MLP-social-GRU.

Predictive Experience Replay for Continual Visual Control and Forecasting

Continual Learning with Deep Generative Replay

An Enhanced Representation Method for Pedestrian Trajectory Prediction Based on Adaptive GCN

Continual Predictive Learning from Videos

BR-GAN: A Pedestrian Trajectory Prediction Model Combined With Behavior Recognition

Pedestrian Trajectory Prediction with Learning-based Approaches: A Comparative Study.

GSGFormer: Generative Social Graph Transformer for Multimodal Pedestrian Trajectory Prediction

Continual Learning of Personalized Generative Face Models with Experience Replay

Multi-Stream Representation Learning for Pedestrian Trajectory Prediction

DDGR: Continual Learning with Deep Diffusion-based Generative Replay.

Socially-Informed Reconstruction for Pedestrian Trajectory Forecasting