Object Tracking Methods
Object tracking is what it sounds like: the process of keeping track of an object as it moves across the screen. This is distinct from object detection in that the program must not only detect a particular object, but also recognize it as the same object across multiple frames of the video. For example, if a car is moving across the road, object tracking involves the program tagging that car with a unique identification as it moves across the screen. The same should be true with multiple objects, where each object consistently has the same ID.
So how exactly does object tracking work? In essence, object tracking comes down to comparing features of objects in adjacent frames, such as proximity, size, and the actual image inside each bounding box.
There are several methods to implement object tracking. I'll give a couple examples. The first is built-in object tracking methods of OpenCV, an open source library for C++ and Python used for processing videos frame by frame. These object tracking methods are convenient, and there are eight different trackers for cases specific to the circumstance. However, the OpenCV method does not provide the same low level flexibility with tracking as can be obtained with writing object tracking from scratch.
Another simple example of object tracking is the Centroid Algorithm, which is relatively simple to implement. The input for this algorithm are bounding boxes from an object detection algorithm. First, for each bounding box, the centroid is calculated (the point in the middle of the rectangle). Then, for one frame to the next, the Euclidean distance from the centroid of every bounding box in the first frame to the centroid of every bounding box in the next frame is calculated. The bounding box corresponding to the closest centroid in the second frame to a particular bounding box in the first is recognized as the same object as in the first frame. Thus, centroid tracking can be used to identify a unique object and keep a record of its location as it moves across the screen.