Social-DualCVAE: Multimodal Trajectory Forecasting Based on Social Interactions Pattern Aware and Dual Conditional Variational Auto-Encoder

Pedestrian trajectory forecasting is essential for applications like self-driving cars, autonomous robots, and surveillance. This task is multi-modal, influenced by physical scene interactions and complex social dynamics among pedestrians. Current literature primarily focuses on learning social interaction representations through deep learning, neglecting explicit interaction patterns like following or collision avoidance. Recognizing these patterns is crucial for accurate forecasting but is challenging due to privacy concerns and label scarcity. We introduce a Social-Dual Conditional Variational Autoencoder (Social-DualCVAE) for multi-modal trajectory forecasting, which leverages a generative model conditioned on past trajectories and unsupervised interaction pattern classification. DualCVAE predicts trajectories by estimating latent variables, with a variational bound as the training objective. Our model outperforms state-of-the-art methods on standard trajectory benchmarks.