Why detection alone doesn't give you identity
An object detector processes each frame independently. It has no built-in sense that the person it just found is the same person it found a moment ago — every frame is, from the detector's point of view, the first frame it has ever seen. Tracking is the layer that adds continuity: associating detections across frames so a single object keeps one identity as it moves through the scene.
The lifecycle of a track
A track isn't a single fixed thing — it moves through states as evidence accumulates or lapses:
- New — just proposed, not yet trusted. A single detection isn't enough to commit to; it could be noise.
- Confirmed — sustained evidence over several frames has raised confidence enough to trust the track and, in many systems, to trigger downstream classification or alerting.
- Coasting / lost — the object briefly stopped producing matching detections (occlusion, a missed frame), but the tracker predicts its likely position and keeps the identity alive for a bounded window rather than closing it immediately.
- Closed — the object left the scene, or the coasting window expired without the track being re-supported.
Each transition has its own evidence threshold. Confirming too eagerly creates false tracks from noise; confirming too slowly means real, briefly-uncertain objects get dropped.
Handling occlusion without losing identity
A person walking behind a pillar for half a second is still the same person on the other side. A tracker that closes a track the instant detection support disappears would treat that reappearance as a brand-new object — breaking dwell time counts, duplicate-alerting, and generally misrepresenting what actually happened. The standard fix is a bounded coasting window: the track survives briefly without direct support, using motion prediction to bridge the gap, and re-associates with the next matching detection rather than starting over.
Association: matching detections to existing tracks
Every frame, the tracker has to decide which new detections belong to which existing tracks. Common signals used for this include spatial proximity (an object usually doesn't teleport between frames), motion consistency (does the new position match where the track predicted it would be), and appearance similarity (does this detection look like the object the track has been following). Getting this association right — especially in crowded scenes where several similar objects are close together — is one of the harder, more benchmark-relevant problems in tracking.
Re-identification: the harder, separate problem
Short-gap association (bridging a half-second occlusion) is different from re-identification: recognizing that a person who fully left the frame and returned five minutes later is the same person, or matching someone across two different cameras with no overlapping field of view. Re-ID typically relies on appearance embeddings rather than motion continuity, and it's a meaningfully harder, more specialized capability — not something every tracking system needs or should assume it has.
It's also worth being explicit that tracking, on its own, doesn't require knowing who someone is. A system can maintain continuity — the same person stays the same track for as long as they're present — without any re-identification or biometric matching at all. See Presence & Dwell Detection for what that looks like in practice.
Where Vision Lab fits
Vision Lab treats tracking as a first-class layer rather than a by-product of detection, and runs it entirely on the edge so footage stays on site. It's the engineering foundation behind ManasaView.
Frequently asked questions
What is object tracking?
Object tracking is the process of associating detections across consecutive frames so the same physical object keeps one continuous identity as it moves, rather than being treated as a new object every time it's detected.
What is a track's lifecycle?
A track typically moves through states such as new (just proposed, unconfirmed), confirmed (sustained enough evidence to be trusted), coasting or lost (temporarily unsupported by a detection, e.g. during brief occlusion), and closed (no longer present). Each transition has its own rules for how much evidence is required.
How does tracking handle occlusion?
A well-designed tracker tolerates a bounded gap in detections — a person passing briefly behind a pillar — by predicting where the object likely is and re-associating it when it reappears, rather than immediately closing the track and treating the reappearance as a new object.
What is re-identification (Re-ID)?
Re-identification is matching an object's appearance to recognize it again after a longer gap or a full loss of track — for example, recognizing the same person after they leave and re-enter a camera's field of view. It's a harder, more specialized problem than short-gap association within a single continuous track.
Does tracking require identity in the sense of "who" someone is?
No. Tracking maintains continuity — the same object stays the same track — without needing to know who or what the object specifically is beyond its class. A presence-and-dwell system, for instance, can count and time how long someone stays without ever attempting to identify who they are.
Why do tracks sometimes split or merge incorrectly?
Crowded scenes, similar-looking objects passing close together, and motion regions that fragment or combine (two people merging into one detected blob, then separating) are the classic causes. This is one of the hardest, most benchmark-relevant problems in tracking, especially as scene density increases.
← Back to Knowledge