Accurate pedestrian path prediction using neural networks
Yi Lu Murphey, Chang Liu, Muhammad Arabi Tayyab, Divyendu Narayan · 2017
This paper presents a study on predicting pedestrian path in a short time horizon, e.g. less than 5 seconds. Our study is conducted within the context of pre-collision detection and avoidance between vehicle and pedestrian using only the positioning data transmitted through V2P communications. An important component in a pre-collision detection system is to accurately predict pedestrian positions in a very short future time period. Three methods are presented, a dead reckoning prediction method, a pattern recognition neural network and a time series neural network, both are designed to predict the future position of a pedestrian based on recent movements. An innovative feature extraction method has been developed for generating feature vectors that are invariant to trip location and prediction time, which are important for training a pattern recognition neural network. All three methods are evaluated on trip data recorded from two pedestrians using a Smartphone application software.