LAUSR

Predicting pedestrian trajectories is crucial for human-interactive systems. This task is compounded by the inherently multimodal nature of human motions and complex external and internal factors such as social interactions and intentions. In this letter, we propose a dynamic attention-based CVAE-GAN method that simultaneously models time-varying social interactions and human intentions while generating multimodal trajectory predictions. Herein, CVAE is used to account for the multimodality of trajectories explicitly, and GAN is used for adversarial training. A spatio-temporal graph with dynamic attention is leveraged to encode pedestrian interactions, which are further fed into a recurrent neural network to capture the temporal dependencies of evolving patterns. Moreover, we aggregate all potential sub-goals from current position to the destination with another dynamic attention module to shape the predicted trajectories. Experiments on widely used ETH and UCY benchmarks will verify the effectiveness of the proposed method. In particular, the proposed method could outperform the state-of-the-art by approximately 10% in terms of prediction errors and 40% considering collision avoidance.