Detecting objects frame by frame produces a fresh list of boxes in every frame, with no memory between them. Frame 1 might report a box at the left edge, frame 2 a box in the middle, frame 3 a box at the right — and nothing in those outputs says these are the same car. Tracking adds that link.
The idea is to maintain a small set of active tracks, each with a remembered position and appearance. When a new frame arrives, each detection is matched to the track it most plausibly continues, based on how far it moved and how similar it looks. Matched tracks are updated; unmatched detections may start new tracks; tracks that go unmatched for too long are retired. The result is a trajectory per object rather than an unrelated list per frame.
This matters because detection alone cannot answer questions about identity. If two people cross paths, a frame-by-frame detector may swap their boxes, and any count of how many people entered a door becomes unreliable. Tracking keeps the identities attached to the right bodies through the crossing, and it also fills brief gaps when a detector misses an object for a frame or two.