Data Association

Data association studies how to decide which detections or measurements belong to the same object, track, or landmark, enabling reliable tracking, mapping, and multi-view reconstruction. Common approaches use motion and appearance costs with Hungarian assignment, probabilistic methods such as JPDA, clustering, and joint estimation; newer systems also use graph neural networks, Sinkhorn-based matching, and learned track states. A central challenge is handling missed detections, occlusion, similar-looking objects, and object births or deaths without losing identity or becoming computationally expensive. An ongoing tension is that better association does not always improve tracking as much as stronger detection, and pairwise ranking metrics may not reflect the quality of globally constrained assignments.