Introduction
A detector looks at one image at a time. In a video, you usually want more: not just "there are 5 people in this frame", but "this is the same person who entered 10 seconds ago, and here's the path they walked". That's multi-object tracking: giving each object a consistent ID across frames.
Tracking is behind counting cars on a road, analysing players in sports broadcasts, following cells in microscopy videos, and helping self-driving cars predict where pedestrians are going.
Think of watching a busy train platform. You don't re-identify every person from scratch each second. You remember roughly where each person was and how they were moving, so you expect them to be near that spot a moment later. You also remember what they look like (red coat, blue backpack), which helps when two people cross paths. Tracking algorithms use exactly those two clues: motion and appearance.