StreamTrack: real-time meta-detector for streaming perception in full-speed domain driving scenarios

Streaming perception is a crucial task in the field of autonomous driving, which aims to eliminate the inconsistency between the perception results and the real environment due to the delay. In high-speed driving scenarios, the inconsistency becomes larger. Previous research has ignored the study of streaming perception in high-speed driving scenarios and the robustness of the model to object’s speed. To fill this gap, we first define the full-speed domain streaming perception problem and construct a real-time meta-detector, StreamTrack. Second, to perform motion trend extraction, Swift Multi-Cost Tracker (SMCT) is proposed for fast and accurate data association. Meanwhile, the Direct-Decoupled Prediction Head (DDPH) is introduced for predicting future locations. Furthermore, we introduce the Uniform Motion Prior Loss (UMPL), which ensures stable learning of the model for rapidly moving objects. Compared with the strong baseline, our model improves the SAsAP (Speed-Adaptive steaming Average Precision) by 15.46 %. Extensive experiments show that our approach achieves state-of-the-art performance in the full-speed domain streaming perception task.