Trajectory prediction of cyclist based on spatial‐temporal multi‐graph network in crowded scenarios

Abstract

Cyclist trajectory prediction is an essential task in autonomous driving and surveillance systems. This task is challenging due to that the bicycles go much faster than the pedestrians and a minor prediction error could lead to a severe deviation in the actual path. Existing cyclist trajectory prediction models usually employ the social pooling mechanism to depict the mutual interactions between targets. They ignore that the pooling operation is leaky in information. Moreover, they prefer to use the recurrent architecture to capture the time-varying features, which is not efficient in computation and parameter learning. To address these issues, a spatial-temporal multi-graph module which employs the topology of graphs to represent social interactions and design multi-kernel functions to depict the social attributes from various aspects is proposed. Instead of the recurrent architecture, a temporal convolution to forecast the future paths is introduced. Experimental results on real-world datasets demonstrate its superior performance against state-of-the-art baselines. It reduces 9% prediction error when compared to recurrent neural network based models and is more effective in crowded scenarios.