A blob representation for tracking robust to merging and fragmentation

This paper describes a new blob representation for tracking that is robust to merging and fragmentation of blobs. The traditional centroid-based blob representation is enhanced by utilizing “central points”, which are 2D local maxima of the distance transform of the blob. We demonstrate the stability of these local features to gross shape changes caused from blob interactions during tracking. A nearest-neighbor blob tracking algorithm is described that reveals the robustness of the proposed representation on standard tracking datasets. Quantitative results show that this representation allows nearest-neighbor blob tracking to achieve robustness comparable to the use of appearance features for data association during blob interactions.