基于跨摄像头的车辆在线跟踪关键技术研究

DSESHADE: Research on Key Technologies for Vehicle Online Tracking Based on Cross-Camera Association

摘要: 多摄像头多目标车辆跟踪(Multi-Target Multi-Camera Vehicle Tracking, MTMCT)由于车辆外观变化、视角差异以及跨摄像头域分布偏移等问题，仍然面临较大挑战。车辆重识别(Vehicle re-identification, ReID)作为跨摄像头轨迹关联的关键技术，在现有数据集约束下，现有方法主要依赖外观特征建模，导致模型鲁棒性与跨域泛化能力受限。针对上述问题，本文构建了大规模多维标注综合数据集VeCamX，通过融合VeRi-776丰富的语义属性信息与AIC22复杂的时空轨迹特征，实现车辆身份、属性、轨迹及姿态关键点等多层次标注统一，为在线车辆跟踪 提供综合性评测基准。 在此基础上，提出一种统一优化框架，融合姿态感知模块(Pose-Aware Module, PAM)、无数据辅助模块(Metadata-Assisted Module, MAM)以及搜索剪枝优化策 略(Search-and-Pruning, SnP)。其中，PAM用于学习车辆目标的视角不变结构特征；MAM 整合车辆类型、 晶牌、 颜色以及时空轨迹等无数据信息；SnP通过构建均衡且高判别性的训练样本集提升模型泛化能力。 在VeRi-776、 AIC22及VeCamX数据集上的实验结果表明，所提出框 架能够稳定提升IDF1 和MOTA 等评价指标，相较现有先进方法表现出更强的鲁棒性与跨场景泛化能力。

Abstract: Multi-Target Multi-Camera vehicle Tracking (MTMCT) remains challenging due to severe appearance variations, viewpoint shifts, and domain discrepancies among cameras. Notably, vehicle re-identification (ReID) plays a crucial role in trajectory association for MTMCT, yet constrained by the limitations of existing datasets, cur- rent ReID-based approaches rely mainly on appearance cues, which compromises their robustness and cross-domain generalization. To address these issues, we in- troduce a large-scale, multi-dimensionally annotated dataset named VeCamX, which integrates the semantic richness of VeRi-776 and the spatio-temporal complexity of AIC22. Specifically, VeCamX unifies annotations for vehicle identities, attributes, trajectories, and poses keypoints, thereby providing a comprehensive benchmark for online vehicle tracking.Building upon VeCamX’s multi-dimensional annotations, we propose a unified optimization framework that combines Pose-Aware Module (PAM), Metadata-Assisted Module (MAM), and Search-and-Pruning optimization strategy (SnP). Specifically, PAM captures viewpoint-invariant structural features of vehicle targets; MAM acquires target-specific metadata including vehicle type, brand, color, and spatio-temporal trajectories; and SnP supplies balanced and highly discrimina- tive data inputs for model training, thereby enhancing the model’s generalization performance. Experiments on VeRi-776, AIC22, and VeCamX show that the pro- posed framework consistently improves IDF1 and MOTA over state-of-the-art meth- ods, achieving strong robustness and generalization in complex real-world scenarios.