PECGAN: Endpoint Conditioned Trajectory Prediction via Generative Adversarial Network

Pedestrian trajectory prediction is a key research topic in the field of computer vision and has been widely used in practical applications, such as robot navigation and autonomous driving. Previous studies predict the future trajectory by decoding the learned motion feature via a self-recurrent architecture, which leads to a significant prediction deviation of the endpoint. Therefore, we propose Predicted Endpoint Conditioned Generative Adversarial Network (PECGAN) to predict the future trajectory without significant endpoint deviations. In our model, endpoint prediction is the primary goal which is accomplished through a conditional variables autoencoder. The estimated endpoints, coupled with past trajectories are encoded as the motion feature, and refined by a social interaction module which adopts the self-attention mechanism for message passing. The refined motion features infer the intermediate trajectory more accurately. Experimental results demonstrate that PECGAN can generate a realistic and diverse set of trajectories that respect physical constraints. Our proposed model improves state-of-the-art performance on the Stanford Drone Dataset benchmark and the ETH-UCY benchmark.