Synchronous Bi-Directional Pedestrian Trajectory Prediction with Error Compensation
Pedestrian trajectory prediction as an essential part of reasoning human motion behaviors, has been deployed in a number of vision applications, such as autonomous navigation and video surveillance. Most existing methods adopt autoregressive frameworks to forecast the future trajectory, where the trajectory is iteratively generated based on the previous outputs. Such a process will suffer from large accumulated errors over the long-term forecast horizon. To address this issue, in this paper, we propose a Synchronous Bi-Directional framework (SBD) with error compensation for pedestrian trajectory prediction, which can greatly alleviate the error accumulation during prediction. Specifically, we first develop a bi-directional trajectory prediction mechanism, and force the predicting procedures for two opposite directions to be synchronous through a shared motion characteristic. Different from previous works, the mutual constraints inherent to our framework from the synchronous opposite-predictions can significantly prevent the error accumulation. In order to reduce the possible prediction error in each timestep, we further devise an error compensation network to model and compensate for the positional deviation between the ground-truth and the predicted trajectory, thus improving the prediction accuracy of our scheme. Experiments conducted on the Stanford Drone dataset and the ETH-UCY dataset show that our method achieves much better results than existing algorithms. Particularly, by resorting to our alleviation methodology for the error accumulation, our scheme exhibits superior performance in the long-term pedestrian trajectory prediction.