Adaptive human silhouette extraction with chromatic distortion and contour tracking
Jamie Schiel, Richard A. Green · 2013
This paper proposes a method to increase the accuracy of human silhouette extraction using background subtraction. Most silhouette extraction algorithms identify shadows when they should not, fail when illumination changes and absorb stationary objects into the background model. An algorithm is proposed that separates the chromatic and brightness changes in a video to clearly differentiate foreground features from background. Large chromatic differences between foreground and background data are used to overcome the effects of illumination changes and differences in brightness are used to overcome shadowing. Chromatic and brightness data collected over 10 frames in a video showed that chromatic distortions in the range (40, 65) indicate foreground objects and brightness distortions below 0.6 indicate shadows. Objects were prevented from disappearing into the background model by tracking the silhouette contours and adaptively updating the background. Experiments showed that silhouette regions should have a background model size of 30 frames while non-silhouette regions should have 5 frames. The proposed method successfully removed shadows and accurately output contours of human silhouettes under varying lighting conditions.