Tracking-Based Semi-Supervised Learning using Stationary Video

This paper deal addresses the semi-supervised problem of tracking and recognizing objects in videos taken with stationary cameras. Building on work on Stanford’s autonomous vehicle using laser range finders to solve the same problem, we aim to develop accurate methods for classifying objects without the additional benefit of 3D laser scans. We set out with three main goals, each building on the previous ones. The first is to perform background subtraction to remove all background objects (those objects that are stationary in the frame of the camera). The second is to track the foreground objects through every frame of the video. Finally, the third goal is to use semi-supervised methods to classify tracked foreground objects. A successful semi-supervised approach will greatly reduce the amount of training data needed for many classification problems.