Kernel-Based Tracking of Geometric Templates

Robust tracking of textured templates in video sequences is an important basic component of many computer vision applications such as surveillance, scene analysis, and shape acquisition. It is also useful in augmented reality and video editing pipeline, to enable texture replacement in existing video sequences. In our work, we extend the mean-shift color tracking approach, and arrive to a histogram-based template alignment algorithm which can handle noisy video sequences.

The basic idea of the method is to replace a single color value at a pixel by a color distribution in its small neighborhood. Following the framework of kernel-based methods, we arrive at a smooth functional to optimize. Combined with local illumination model, this gives a robust template tracking procedure.

In the talk, I will start from a brief overview of my current research projects in geometry processing, and then proceed with motivation, details, and examples from the kernel-based video tracking project.