Graph cut needs user interaction on image segmentation, which comes up with that threshold value method, gets the objects by cutting rectangle and then we remaps to image to finish the interaction by laser radar. GrabCut algorithm is sensitive to local noise, and it is time-consuming. In addition, the edge extraction is not ideal under complex background, so an improved GrabCut algorithm is put forward to adapt background automatically in the determined background. The proposed algorithm chooses probable foreground neighboring pixels as background pixels to make background become simple. It is applicable to the case when foreground pixels account for low proportion in the whole image pixels and the background pixels are relatively simple around the foreground. Experimental results show that error rate of the proposed algorithm is reduced and the efficiency is improved in comparison with GrabCut algorithm after reducing nodes number in the graph. In addition, the edge extraction is more complete and smooth.