Fast Parallel Object Tracking

The task of object tracking is the following: given a video and an identified object (usually given by a bounding box in the first frame), track the position of the object over the frames of the video. Difficulties in this process include occlusion or changes in orientation. The algorithm by Grabner, Grabner, and Bischof [3] is robust against these changes by using a large number of “weak classifiers” and “selectors” and is illustrated in Figure 1. The main building block of this system is the “weak classifier.” These are relatively simple objects that classify a group of pixels as the target object positively or negatively. The most important part is that they must do better than 50-50 guessing (a relatively loose condition). This heavy relaxation allows these types of classifiers to be computed very quickly. Selectors have the job of boosting the accuracy far beyond 50-50 guessing. By choosing the most accurate weak classifiers, selectors are much more likely to accurately pinpoint the location of the object. Further, combining the outputs of the selectors creates a confidence map that accurately and precisely locates the object. The input to the algorithm is a new frame. This is then preprocessed and then passed to the weak classifiers to create individual confidence maps. These confidence maps are then chosen by the selectors based on error rate and linearly combined to create a confidence map for the strong classifier. This can then be translated to a coordinate which is the output of the algorithm. We then treat the coordinate to be the truth and update all weak classifiers to better recognize the object in the future. I chose to use a weak classifier based on Haarlike features. In essence, these take the weighted sum of rectangles of pixels and use a threshold to determine classification. The upside of this type of classifier is that it can be computed