Fast Automatic Video Object Segmentation for Content-Based Applications

Abstract

An algorithm has been devised for fast, fully automatic and reliable object segmentation from live video for scenarios with static camera. The contributions in this chapter include methods for: (a) adaptive determination of the threshold for change detection; (b) robust stationary background reference frame generation, which when used in change detection can reduce segmentation fault rate and solve the problems of dis-occluded objects appearing as part of segmented moving objects; (c) adaptive reference frame selection to improve segmentation results; and (d) spatial refinement of modified change detection mask by incorporating information from edges, gradients and motion to improve accuracy of segmentation contours. The algorithm is capable of segmenting multiple objects at a speed of 12 QCIF frames per second with a Pentium-4 2.8GHz personal computer in C coding without resorting to any code optimization. The result shows advantages over related work in terms of both fault rate and processing speed.