Multi-directional geometric enhancement of optical flow approach for moving objects recognition

摘要

A variant of the Lucas-Kanade tracking procedure jointly used with the infinitesimal interactive affine motion approximation of displace- ment, are applied to subsequent camera-acquired images. The outputs are subject to ICA, followed by a variance-based decision criterion, in order to outline the moving object (assumed to have a closed non self-intersecting contour). The output determines the segmentation of the image into mov- ing obstacles and dominant plane.The originality of the approach resides in using a unitary circular Gateaux-derivative treatment of all neighboring directions in the process of preliminary motion-evaluation based on pixel shift decomposition, which allows better detection of significantly mobile clusters of pixels. In order to highlight our results, a comparison was made between the proposed algorithm and some others from literature. The algorithm can be applied to a large class of images and is intended to be an interactive tool in obstacle-detection for single-camera endowed moving robots.