Research / Project
PairGate3D
第一作者 · IEEE-RICAI 2026 已录用
摘要
Online 3D multi-object tracking maintains vehicle identities through frame-wise detection association. In score-weighted association, track scores affect both candidate ranking and, under a fixed matching threshold, the spatial acceptance region, admitting distant and weakly overlapping pairs to assignment. We present PairGate3D, a complete online tracking system with a score-independent geometric gate. Before assignment, the gate rejects a candidate edge when its 3D center distance is large and BEV overlap is low, while retaining the original costs and ranking of all other candidates. The thresholds can adapt to prediction uncertainty or be calibrated on target scenes. Gating improves HOTA across five KITTI Car detector streams, including a +5.757-point gain on PointRCNN. Evaluations on V2V4Real and nuScenes further show how geometric eligibility supports tracking accuracy and identity continuity across datasets.
研究问题
在线三维多目标跟踪根据连续帧的目标检测结果维护车辆轨迹。跟踪器需要判断当前检测框对应哪条已有轨迹，并在目标短暂漏检后继续预测其位置。当错误的检测框被接入轨迹，位置和身份信息都可能受到影响，后续帧中的预测也会随之偏移。
论文关注关联过程中的候选匹配边界。在使用轨迹分数加权的关联代价中，分数会影响候选匹配的排序；当系统还使用固定代价阈值决定候选是否进入分配时，分数也会改变候选在空间上的接受范围。相同中心距离的轨迹与检测框，可能因轨迹分数不同而得到不同的准入结果。
研究的问题是如何在分数加权的候选排序之外，单独约束检测框与轨迹的几何关系。这样可以在进入分配算法之前，检查空间上相距较远且重叠较小的候选组合。