Real-Time Visual Object Discovery for Tracking Using Motion and Depth Data

The ultimate goal of our method is to be able to discover and track any object in any scene with minimal prior knowledge. This semester thesis focuses on a real-time implementation that uses motion and depth as cues to discover objects. More precisely, feature points are detected using the minimum eigenvalue corner detector and calculate the optical flow of these points by applying the Lucas-Kanade method. We then use a graph-based segmentation algorithm to group the detected feature points either with depth or motion similarity. The clustered features are then examined with an objectness measure to output bounding boxes which are likely to contain an object. A mentionable speed improvement has been achieved.