A new integrative approach to time varying image interpretation

This paper presents a global method to process monocular image sequences for mobile robot obstacle detection. We do not aim to achieve a complete scene reconstruction, but only to evaluate the time to collision and surface orientation useful for robot obstacle avoidance. Using a planar facet representation we first calculate formally the velocity field generated by the camera undergoing an arbitrary motion. Then, we propose to interpret this vector flow along image contours, using an integrative approach in conjunction with a projective model, and we introduce a link with image moments. Under the assumption of an approximate axial motion, we show that our integrative approach can be applied as well to estimate the image vector flow without involving any prior feature correspondence. Noise resistance is improved by integration. Experimental reconstruction results on images sequences are given, and their sensitivity to the various perturbations of camera motion is evaluated. Potential applications include visual navigation, obstacle detection, visual servoing, depth from zooming, object recognition.