Vehicle-infrastructure cooperative general object detection through feature flow and differentiable pose-based spatial alignment

Vehicle-infrastructure cooperative perception is critical for advanced autonomous driving but faces challenges: heavy reliance on ground truth (G.T.) vehicle-infrastructure poses for spatial synchronization and limitations of traditional object detection in dynamic scene modeling. This study proposes a differentiable pose-based spatial alignment (DPSA) module and explores occupancy flow output for cooperative perception. The DPSA module eliminates G.T. pose requirements by estimating 3-degree-of-freedom relative poses via feature fusion, spatial average pooling, and fully connected layers, balancing accuracy and practicality. Evaluations on DAIR V2X and V2X-Seq datasets show DPSA outperforms misaligned fusion methods in key metrics, maintains robustness under low-location precision scenarios, and reduces inference time. Notably, occupancy flow demonstrates superior dynamic modeling capability compared to traditional object detection by capturing spatial occupancy changes and motion flows (instead of bounding boxes), enabling high accuracy future occupancy prediction and occlusion resistance via multi-source temporal context. This work advances practical cooperative perception through an efficient synchronization solution and highlights occupancy flow advantages, bridging algorithmic innovation and engineering deployment.