An efficient object segmentation algorithm with dynamic and selective background updating and shadow removal

A dynamic and robust variant of the video object segmentation algorithm due to Horprasert et al. [5] is investigated. The new method overcomes the limitations of the prior art while addressing practical problems in realistic visual surveillance tasks such as highly compressed video data, environmental changes, internal shadows, etc. Based only on 'brightness distortion' metric the thresholds for classifying 'nonbackground' pixels, or shadow / highlight detection, are determined automatically and then updated cyclically. In addition, we also introduce an active contour-based internal hole refilling technique to ensure an object's integrity as well as a detection feedback mechanism to deal with object deposit/removal into/from the scene. Test results on highly compressed video data are illustrated as part of an integrated object tracking system.