When two tracked people overlap, then there is some ambiguity as to which model the box really belongs to. Usually at the initial phase of overlap, the two targets are still visible. Then after a bit, if one person really passes behind another, only the single person is visible. Eventually then, the person re-emerges but is still part of the same detection window. The final phase involves the two people separating and returning to being two distinct track regions.
The appearance information and the re-identification model provide information regarding who is the front-most person (the one occluding the other). Let's see if we can differentiate and identify who is really the target. Come up with a strategy to detect when two target bounding boxes overlap and join to become one. At that point, the track ID of the single target should be set to uncertain, as potentially belonging to the two people (or more if you get a triple person merger; that's tougher). As they merges and split, the area of the bounding box should grow significantly, shrink, then grow again. Create an algorithm that tracks the target ID's area (with some smoothing most likely) and identifies when that area is at a minimum.
Once the minimum is found, that should represent the time point of greatest occlusion. Using the area within the bounding box, apply the appearance model for re-identification and give the object the target ID of the best fit model. Hopefully when the second person re-emerges, the re-identification part matches the person to the proper target model and ID.
Deliverable: Identify one or two scenarios in your videos where this happens. Show the appearance based re-identification model applied to them such that the main ID of the merged targets can be identified.