RECUP Net: RECUrsive Prediction Network for Surrounding Vehicle Trajectory Prediction with Future Trajectory Feedback

Abstract

In order to predict the behavior of human drivers accurately, the autonomous vehicle should be able to understand the reasoning and decision process of motion generation of human drivers. However, most of the conventional prediction methods overlook this and focus on improving the prediction results using the given data, the historical information. In contrast, human drivers not only depend on the historical motion but also consider future predictions when handling interactions with other vehicles. In this paper, we propose a novel recursive RNN encoder-decoder prediction model that takes the initial future prediction results as inputs of second prediction computation. This feedback mechanism can be interpreted as a network sharing, which allows the model to refine or correct the predicted results iteratively. We use two encoders to analyze both of the historical information and future information, and the attention mechanism is employed to interpret interaction. Our experimental results with the NGSIM dataset demonstrate the recursive structure enhances prediction results effectively compare to the baselines based on the ablation study and state-of-the-art methods. Furthermore, we observe that the results improve successively as the model iterates.