Abstract: Collaborative perception realizes joint object detection in complex traffic environments via multi-vehicle information sharing and fusion. Existing approaches mainly rely on single-frame or short-term feature fusion and fail to fully exploit historical information, leading to inconsistent and unstable detection of distant objects across time. To address this limitation, a collaborative perception method based on closed-loop collaborative trajectory sharing was proposed. Built upon intermediate collaborative fusion, trajectory-level historical information sharing and feedback were incorporated to form a two-stage closed-loop architecture. Cascaded matching and Kalman filtering were employed to continuously estimate and maintain multi-frame object states, while a Multi-View Trajectory Fusion module measured geometric consistency among multi-view trajectories using the Hausdorff distance to achieve cross-vehicle trajectory fusion and collaborative refinement, thereby improving temporal stability and spatial robustness. Experimental results on OPV2V and V2V4Real public datasets show that, under an Intersection over Union (IoU) threshold of 0.7, incorporating the proposed method into V2VAM (Vehicle-to-Vehicle Attention Module) and DMSTrack (Differentiable Multi-Sensor Tracking) improves the mean Average Precision (mAP) by 2.7 and 5.7 percentage points, respectively. Ablation studies further confirm the effectiveness and generality of the closed-loop trajectory sharing mechanism.

摘要： 协同感知通过多车信息共享与融合，实现复杂交通环境下的联合目标检测。现有协同感知方法主要依赖单帧或短时特征融合，对历史信息利用不足，导致远距离目标在跨时间观测中难以保持一致且稳定的检测结果。针对上述问题，提出一种基于闭环协同轨迹共享的协同感知方法。该方法在中期协同基础上引入轨迹级历史信息共享、协同与反馈策略，构建两阶段的闭环协同感知体系。其中，动态轨迹跟踪模块通过级联匹配与卡尔曼滤波实现多帧目标状态的持续估计与稳定维护；多视角轨迹融合模块利用Hausdorff距离度量多视角轨迹几何一致性，实现跨车辆轨迹的精准融合与全局协同优化，从而显著提升检测结果的时序稳定性与空间鲁棒性。在OPV2V、V2V4Real公共数据集上的实验结果表明，在交并比(Intersection over Union，IoU)为0.7的条件下，将本文方法引入V2VAM(Vehicle-to-Vehicle Attention Module)与DMSTrack(Differentiable Multi-Sensor Tracking)后，其平均检测精度分别提升了2.7、5.7个百分点。相关消融实验进一步验证了闭环轨迹共享机制在提升协同感知稳定性与全局一致性方面的通用性和有效性。