Improving Interaction-based Vehicle Trajectory Prediction via Handling Sensing Failures

Accurately predicting the future trajectories of surrounding vehicles is critical to the safety and efficiency of autonomous vehicles. Although current trajectory prediction models based on deep learning and vehicle interactions have made substantial progress in accuracy, ensuring the performance of these models in the real-world has become a new challenge due to inevitable sensing failures in running autonomous vehicles. In this paper, we propose a prediction-oriented sensing enhancement framework to improve the performance of existing interaction-based trajectory prediction models on real-world vehicles with sensing failures. The proposed framework consists of an object lifecycle management module and a graph matching model. It focuses on handling sensing failures such as missed detections, false detections, and ID-switch caused by sensor limitations or non-robust perception algorithm. By monitoring the lifecycle of detected objects and re-matching newly appeared and recently disappeared key nodes, the proposed method efficiently reduces the scattered tracklets inputted into the prediction model and thereby improves the prediction performance. Experimental results on the large public dataset nuScenes and real-world AVs demonstrate that our proposed method is effective and practical.