Pedestrian Trajectory Prediction in Crowd Scene Using Deep Neural Networks

Abstract

Pedestrian trajectory tracking prediction is important at intersections to human’s safety and thus requires the designing of intelligent driving systems. Accurate prediction of the pedestrian path is a priority to design a reliable system for tracking the movements of humans in a crowd. There are several techniques proposed to predict the trajectory of pedestrians. Long short-term memory (LSTM) is based on a recurrent neural network. The other one is social LSTM, called social pooling, which combines the human–human interaction model. However, long-range dependency is not properly described in existing approaches, which ignores the semantic information for trajectory tracking. From another point of view, there are many techniques for data collection and processing, such as image processing and extraction features to determine the location, speed, etc. However, changing the lighting conditions and the difference between the daylighting conditions and the night lighting creates many challenges. The study systematically presents trajectory prediction methods using deep learning network architecture including datasets used to evaluate trajectory prediction methods.