MHCPP: A Motion-Based Historical Enhancement Collaborative Perception and Prediction Framework

Multi-agent collaborative perception mitigates the issues of limited perception range and occlusion in single-vehicle perception through shared sensing information, resulting in improved perception performance. However, current collaborative strategies often fail to differentiate moving objects from the background in complex scenarios. This leads to critical object data being obscured by excessive background information during data fusion, complicating effective integration. Additionally, efficiently leveraging historical perception data remains a significant challenge. To address these issues, we propose a parameter-free motion attribute extraction module that captures motion features between consecutive frames, enabling more efficient utilization of historical information to enhance current perception. Specifically, we design two temporal modules to comprehensively explore both ego and collaborators’ historical information. Furthermore, we propose a method that simultaneously considers the relationships between global appearance features and local motion features of shared information, thereby enabling better integration of foreground object data during collaboration. We validate our algorithm through experiments conducted on both simulated and real-world collaborative benchmarks. Both qualitative and quantitative results demonstrate that our proposed MHCPP significantly outperforms state-of-the-art (SOTA) methods in terms of 3D object detection and prediction tasks.