Pedestrian Trajectory Prediction With Recurrent Nerual Network

In complex and crowded scenes,accurate trajectory prediction of pedestrians is a critical task in areas such as autonomous driving and robot navigation.Pedestrian trajectory in realistic scenes is affected by many factors,The previous pedestrian trajectory prediction methods based on social force model can't solve this problem well by handcraft rules.Recent studies based on recurrent neural network model(RNN)provide new ideas for the pedestrian trajectory prediction.Our work proposes two pedestrian trajectory prediction models based on the Long Short Term Memory network(LSTM).Combining the social force model and LSTM,we propose the DSV-LSTM model for pedestrian trajectory prediction,which a special DSV-Pooling layer is added to.The DSV-Pooling layer models the interactions among pedestrian trajectories based on factors such as relative speed,relative distance,view range and view angle.Then model combines the influence information of the DSV-Pooling layer with the hidden states of trajectory LSTMs to represent the pedestrian interaction information.The DSV-LSTM model simulates the interactions and pedestrian motion decision process in complex and crowded scenes very well.And we also propose a Dual-Attention model combined attention mechanism and LSTM.This model uses the soft attention mechanism to model the impact of pedestrian history trajectory information on future trajectory.It also uses the self attention mechanism to consider the interactions among pedestrians.Dual-Attention model does not require hand-craft rules and corresponding parameters.And it's a data-driven and end-to-end pedestraian trajectory prediction model.We validate the DSV-LSTM model and Dual-Attention model on two public pedestrian trajectory datasets UCY and ETH and the two models achieve state-of-the-art results.