Precise Extraction of Moving Objects in Video Stream

Abstract

We propose an automatic and precise moving-object extraction method for use in video streams that can also be used for 3-D system applications. The method generates a statistical model for each pixel using several frames, and then uses it to generate trimap images. After manually initializing a frame, unknown regions are automatically determined either background or foreground for the rest of frames. The key technology proposed is an adaptive training scheme, which estimates detection thresholds locally through the algorithm, followed by matting approaches using an iterative process and weighted statistical distance minimization. Experiments demonstrate outperformance of our method for both indoor and outdoor video streams, and also for 3-D modeling and representation.