Impact of Posture and Social Features on Pedestrian Road-Crossing Trajectory Prediction

A comparative experiment on machine learning (ML)-based pedestrian trajectory prediction was conducted in this study. The study was aimed at improving functional capabilities, where pedestrian safety is the key, such as in relation to intelligent transportation, autonomous vehicles, and automated-guided vehicles. In this article, the effects of integrating social and posture features into pedestrian trajectory prediction are explored. The experimental results show that incorporating social and posture features can improve the accuracy of pedestrian trajectory prediction. Furthermore, various deep learning (DL) models were adopted and compared to deal with these different types of features. Our work shows that a DL approach can handle complex features better than conventional ML. To enable the neural network process different types of data, four types of networks are designed, namely, C-network, S-network, P-network, and SP-network, which are a 2-D convolutional neural network (CNN) with only coordinates, 2-D CNN with the social feature, 2-D CNN with posture features, and 3-D CNN with social and posture features, respectively. The angle mean square errors (MSEs) of SP-network are 4.99, 1.31, and 0.05 less than those of C-network, S-network, and P-network, respectively. The scalar MSEs of SP-network are 0.204, 0.024, and 0.154 less than those of C-network, S-network, and P-network, respectively. Moreover, feature visualization and importance are discussed. It is concluded that the posture features make it possible to separate features with different scalars and angles. It is also shown that the most important posture feature is the edge of the shoulders.