Pedestrian Trajectory Prediction in Graph Representation Using Convolutional Neural Networks

Predicting the future trajectories of pedestrian in real-world contexts, including video surveillance, self-driving, and robotic systems, is a challenging task due of different trajectory patterns. In this task there are two primary problems: (1) advanced interaction modeling among pedestrians and (2) the specific motion pattern extraction. Pedestrian trajectories are affected not only by another pedestrian, but also by interactions with the environment. To obtain the interactions of pedestrian movements and the active change trend of the environment, there are two components in the proposed method: encoder with a spatial graph neural network for interaction modeling and decoder with a temporal graph neural network for motion pattern extraction. The investigated approach is more compact and efficient than the based method, with a reduced variable size and better accuracy. Furthermore, utilizing two publicly available datasets (ETH and UCY), our model achieves better experimental results considering final displacement error (FDE) and average displacement error (ADE) metrics and predicts more socially appropriate trajectories.