Re-identification

In computer vision based multi object tracking, re-identification (ReID) can be used for addressing the issue of objects temporary being occluded or leaving the scene. Objects being tracked can be temporarily hidden by other objects or move out of the camera's view, making consistent tracking challenging. Additionally, variations in pose, orientation, and lighting conditions from one frame to another can further complicate the tracking.

When an object temporary disappears and then reappears in a new frame, the tracker may fail to recognize it and start tracking it as a new object initiating a new track, leading to errors and inconsistencies. This misidentification disrupts the continuity of tracking information.

ReID aims to resolve this by matching the features of the reappearing object with those of the previously tracked objects, regardless of changes in location, orientation, or lighting. This approach improves the trackers ability to maintain consistent tracking information for the object throughout the video.

This image illustrates how a re-identification might happen in a scene, and how a tracker may indicate the result of the ReID. We observe that the human track first has id 1 then, as the object leaves the scene and comes back, the tracker then initiates a new track with id 2 for the same human. After some time however, ReID kicks in and a rename is triggered saying that the track with id 2 is actually tracking the same object as the track with id 1. This indicates the track with id 2 may be incorporated into the track with id 1, the track with id 2 is terminated and tracking of the object will be resumed using the track with id 1.