Robust Construction of Spatial-Temporal Scene Graph Considering Perception Failures for Autonomous Driving

Spatial-temporal scene graphs (STSG) are emerging for motion prediction in autonomous driving. Existing work focuses on the graph structure and corresponding graph neural network models, ignoring the challenge of constructing STSG on real-world autonomous vehicles. In this paper, we propose a method for robustly constructing STSG against perception failures that may occur in real-running vehicles. We first propose an object-oriented lifecycle management module to identify abnormal nodes by scoring to deal with the possible missed detection and false detection in perception. Then we employ Kalman filter to predict the state of the missing nodes to complete the lost information, and develop a novel bipartite graph matching strategy based on the Kuhn-Munkres algorithm to re-match the abnormal nodes. Experimental results on public datasets show that our proposed method can effectively correct possible errors in raw perception results, thereby improving the stability and reliability of the constructed STSG.