Re‐Identification‐Based Models for Multiple Object Tracking

The multiple object tracking problem is formulated using the assumption of linearity and independence of the objects movement. Factorization of the posterior distribution of objects' parameters proves that the original problem is equivalent to the procedure containing two steps: track prediction and assignment of objects. To achieve the stability of assignments, an assignment cost is developed based on the re‐identification of the candidate descriptor from the track history. Given that track measurements are not equal in terms of usefulness for re‐identification, the technique of track descriptor pre‐filtering is introduced in the current study. Known quality assessment methods and an alternative detector‐based approach are considered. Computational experiments were conducted on multi‐object tracking (MOT20‐01, MOT20‐02) datasets. The results showed the computational efficiency of the re‐identification‐based methods and also the increased stability of tracking in complex scenarios.