A Spatio-Temporal 3d Convolutional Neural Network to Predict Pedestriantrajectory in Heterogeneous Traffic

Predicting pedestrian trajectory is a critical ability for an autonomous driving vehicle while driving through tvraffic. The pedestrian trajectory is governed by the spatial and temporal relationship of the environment, i.e., the physical space and heterogeneous traffic agents. It is important to model these relationships between the pedestrians, physical space, and heterogeneous traffic agents to predict the pedestrian trajectory. The existing approaches mainly use LSTM based approaches to exploit the local spatial correlations, including pedestrian-pedestrian interactions and pedestrian-scene interactions among the spatially adjacent regions. However, the non-Euclidean correlations from the global scene context and temporal correlations, including pedestrian heterogeneous agent’s interactions and pedestrian-space interactions among the possibly distant regions are critical for forecasting pedestrian trajectory. This study proposes a spatial and temporal correlations framework that is simultaneously based on 2D and 3D convolutions for trajectory forecast. The 2D convolutions encode the spatial features, whereas the 3D convolutions encode the temporal features. The attention module captures the non-Euclidean correlation among the remote regions. Finally, the spatiotemporal relations that significantly influence the trajectory of the subject pedestrian are used to forecast the subject pedestrian trajectory. This model is evaluated on public benchmark datasets, and this investigation demonstrates that the proposed network outperforms state-of-the-art methods.