Real-Time Multi-Resolution Blob Tracking

This paper introduces a new real-time blob tracking algorithm. Segmentation is the first step in many video analysis approaches. A number of successful segmentation techniques extract regions of interest, or blobs, in successive frames. The problem addressed here is that of establishing temporal relationships between blobs, without the use of domain-specific information. These relationships can then be analysed at a higher semantic level. The proposed algorithm maintains a multi-resolution tracking graph that encodes, at each resolution, the temporal relationships of the blobs detected in successive frames. As blob displacement in image space is reduced in lower resolution levels, a coarse-to-fine correspondence hypotheses generation, propagation and refinement approach allows to track not only large, slow blobs but also small, fast blobs. Tracking performance is illustrated on various simple application scenarios using a real-time implementation of an integrated segmentation and tracking system. Blob tracking results are demonstrated on standard video surveillance datasets, as well as real-time ball (and player) tracking results in professional tennis and racquetball videos.