The Taub Faculty of Computer Science Events and Talks

Object tracking is one of the basic and difficult aspects of computer vision. It is generally a prerequisite for other computer vision tasks such as object recognition, as well as a goal by itself in applications such as video surveillance. We address the problem of visual tracking in a general context using two approaches. Our first approach is the combination of multiple trackers that use different features and thus have different failure modes. We propose a general framework for combining visual trackers that propagate filtering distributions over time. The individual trackers may propagate the filtering distributions either explicitly, e.g., via Kalman filtering, or by using sample-sets of the distributions, via particle filtering. The proposed framework enables the combination of trackers of different state spaces, and in many cases it allows treating the individual trackers nearly as "black boxes." Another benefit of the proposed framework is that it may be applied as is to the combination of trackers that track different, albeit related, targets. Our second approach consists of the employment of basic, low-level visual characteristics, which are typically valid. We propose a tracker of object bitmaps without using any prior information about the target or the scene. The tracker relies only on basic visual characteristics, which makes it applicable in a very general context. In addition, we propose a novel kernel-based tracker that exploits the constancy of color and the presence of color edges along the target boundary. This tracker uses these two visual cues to affinely transform the kernel over time. In this talk, I will also present our extension of the mean shift tracker to using multiple reference color histograms.