Why trackers are slow
A common way to track objects is to detect them in every frame, crop each detection, and run a second network on the crops to produce an appearance embedding. Matching embeddings across frames keeps identities stable when people cross paths. It works, but the second network scales with the number of objects. A busy aisle with 30 people can cost more than the detector itself.
On a device with an 8 ms budget, a cost that grows with the crowd is a cost you cannot plan for.
Reusing what the detector already knows
The detector has already looked at every object at several scales. Its intermediate features contain most of what an appearance network would compute. Depth-Track samples those features at each box, projects them into a small embedding, and matches them with motion predicted from the previous frames.
- No crops and no second backbone, so cost stays nearly flat with object count.
- Embeddings are 64 values, small enough to keep a long identity history on chip.
- Motion and appearance are weighted by how confident each one is in the current frame.
Cost and quality
Depth-Track adds 1.2 ms on a Hexagon NPU with 30 tracked objects and 1.4 ms on an Ambarella CV72S. It scores 63.4 HOTA on MOT17, within two points of trackers that run a separate re-identification network at several times the cost.
The cheapest network is the one you already ran.
Using it
Depth-Track ships as an attachment to any Depth detector. Add it to a graph in Parallax Studio or in code, and every event in the stream carries a stable track identifier. Identities survive short occlusions of up to two seconds by default, which can be tuned per camera.