Automatic foreground object segmentation from multi-view images

In order to segment foreground objects from multi-view images automatically, an object extraction algorithm is proposed from multi-view images. Firstly, straight lines in epipolar plane image (EPI) are detected by an improved randomized Hough transform, and followed by a statistical analysis of the slopes of EPI lines. Then, according to the depth of the object under segmentation, the corresponding EPI lines are mapped back to the original image space to get an initial contour. After that, an edge growing method is used to reduce the gaps between broken edge segments. Finally, an edge linking method is adopted to obtain a closed contour curve. Experimental results show that, compared with the active contour segmentation methods based on level set, our method can extract objects from complicated scene more accurately and more efficiently.