Object detection and tracking using a parts-based approach

Daniel Clark

Note: imported from RIT’s Digital Media Library running on DSpace to RIT Scholar Works. Physical copy available through RIT's The Wallace Library at: TA1634 .C54 2005


One of the main goals of artificial intelligence is to allow computers to understand the world around them. As humans we extract a large amount of knowledge about the world from our visual perception, and the field of computer vision is determined to give computers access to this same wealth of knowledge. One of the fundamental steps in understanding the world is finding specific objects within our field of view, and the related task of following these objects as they move. In this thesis the Implicit Shape Model algorithm, a local feature-based object detection algorithm, is implemented and used to develop an appearance model and object tracking algorithm based on it. This algorithm is very robust to intraclass variation, and can successfully track objects when both occlusion and non-stationary backgrounds are present. The usefulness of the proposed appearance model is analyzed, and results of the algorithm on real video sequences are presented. Several enhancements to the method are also proposed, and performance in terms of recall and precision is analyzed.