Video Tracking Based on Spectral Graph Matching
In this talk, we address the problem of tracking an arbitrary object in a video given its location in the first frame and no other information. In our work we combine long term motion modeling of trajectories with the requirement to generate object labels at each patch site. Our tracking system unifies color, spatial and temporal parameters in a graph in order to solve the problem using Spectral Graph Matching. KL-divergence is used in order to model the color distance, while adjacency between different segments reflects the spatial parameter. The temporal parameter is a combination of previous frame results and Multiple Instance Learning.
This talk introduces a robust tracking system without any parameter tuning during the process and dealing with a tracking failure.
*A Research thesis under the supervision of Dr. Yosi Keller