Crossing-Road Pedestrian Trajectory Prediction via Encoder-Decoder LSTM

In urban road scenarios with coexistence of vehicles and pedestrians, the ability of predicting pedestrians' future position is essential for the intelligent vehicle to avoid potential collision risk and make reasonable path planning. For vehicles and pedestrians, the behaviors and states of both sides will affect each other to make their judgments of "right of way". However, most of the previous works have ignored the interaction characteristic of traffic participants in the pedestrian trajectory prediction task. which could hardly describe the interaction scenario. We proposed a novel network architecture based on the encoder-decoder Long Short-Term Memory (LSTM) network. A double-channel encoder is designed to extract the state streams from both vehicle trajectory and pedestrian trajectory. Then the state fusion is implemented in the decoder to generate the future trajectory of pedestrian. In experiments, our method has been compared with both Dynamical Motion Models based method and data-driven based method. The results verified the effectiveness of our method especially on the Daimler dataset and a new established dataset VPI. The results verified the effectiveness of our method especially in long term prediction.