Data Association
Data association is the problem of deciding which sensor observation belongs to which object when several are being tracked at once, amid noise and false alarms. We explain why it is hard, the classic methods (nearest neighbor, JPDA, multiple hypothesis tracking), its relation to the Kalman filter and its modern version with deep learning.
Data association is the problem of deciding which observation or measurement from a sensor belongs to which object, when several objects are being tracked at once. It is a central piece of multi-target tracking: at each instant, detections arrive—radar echoes, boxes in a video—and they must be correctly distributed among the trajectories being maintained.
Why it is hard
The difficulty stems from uncertainty. The sensor does not say which object each measurement comes from; there are false alarms (noise that looks like an object), missed detections (objects not seen at some instant) and an unknown number of objects that also changes over time. When two objects cross or come close, telling which is which becomes especially delicate.
The classic methods
Several strategies exist. The simplest is nearest neighbor, which assigns to each trajectory the closest measurement; its global version chooses the globally most probable set of pairings. Joint probabilistic data association (JPDA) does not decide all at once: it weights all possible associations by their probability and updates with that mixture. And multiple hypothesis tracking (MHT), formulated by Donald Reid in 1979, keeps several association hypotheses open over time and resolves them with later evidence.
Its relation to the Kalman filter
Association does not work alone: it combines with the Kalman filter, which predicts where each object will be and estimates its state. The filter's prediction serves to bound which measurements are plausible for each trajectory, and association decides which of them feeds the update. Filtering and association are the two faces of tracking.
The modern version
In 2026 the tracking-by-detection paradigm dominates: first a detector locates the objects in each frame, and then association links those detections over time. Methods such as SORT (2016) combine the Kalman filter with an optimal assignment, and DeepSORT (2017) adds a visual appearance descriptor, learned with neural networks, so as not to confuse identities after an occlusion. It is a key technology in autonomous driving and sensor fusion.
This article was produced with artificial intelligence under human editorial oversight.
