Abstract

The subject of the presentation is concerned with online tracking of moving foreground objects through an image sequence. As part of the fourth year honours project, relevant parts of the paper ``Learning patterns of activity using real-time tracking`` (Stauffer & Grimson, 2000) were implemented and evaluated. Specifically, an adaptive background subtraction technique, based on pixel-wise Mixtures of Gaussians is used to estimate a multimodal distribution over background pixels. The mixture representation allows repetitive background motion to be absorbed into the background estimate. Foreground objects in the scene become apparent as ``blobs`` using outlier detection on each pixel`s mixture. To track these blobs across several image frames, Kalman filtering is employed. In the context of tracking multiple objects simultaneously, determining the association between detected blobs and tracked objects is a non-trivial problem, not fully described by Stauffer and Grimson. The association problem is formulated as an instance of the linear assignment problem and solved using existing algorithms. The results indicate that multiple objects can be tracked in simple scenes reasonably well.