Collision-Free LSTM for Human Trajectory Prediction.

Pedestrians have an intuitive ability for navigation to avoid obstacles and nearby pedestrians. If we want to predict future positions of a pedestrian, we should know how the pedestrian adjust his direction to avoid collisions. In this work, we present a simple and effective framework for human trajectory prediction to generate the future sequence based on pedestrian past positions. The method, called Collision-Free LSTM, extends the classical LSTM by adding Repulsion pooling layer to share hidden-states of neighboring pedestrians. The model can learn both the temporal information of trajectories and the interactions between pedestrians, which is in contrast to traditional methods using hand-crafted features such as Social forces. The experiments results on two public datasets show that our model can achieve state-of-the-art performance with assessment metrics.