Detection has no memory, tracking is about identity over time
Run an object detector on a video by itself, frame by frame, and it treats every single frame as a brand new, unrelated image. Frame 10 might detect 'person at this location' and frame 11 might detect 'person at a slightly different location,' but the detector has no built-in concept that these are the same person. Each frame is detected completely from scratch, with zero memory of what came before.
Tracking sits on top of detection and adds exactly that missing memory. Its job is to assign a consistent identity, often just an ID number, to each object and carry that identity forward across frames, so 'person A' detected in frame 10 is recognized as the same 'person A' in frame 11, frame 12, and so on, rather than being treated as a fresh, unrelated detection every time.