Fast object segmentation from a moving camera

Segmentation of the scene is a fundamental component in computer vision to find regions of interest. Most systems that aspire to run in real-time use a fast segmentation stage that considers the whole image, and then a more costly stage for classification. In this paper we present a novel approach to segment moving objects from images taken with a moving camera. The segmentation algorithm is based on a special representation of optical (low, on which u-disparity is applied. The u-disparity is used to indirectly find and mask out the background flow in the image, by approximating it with a quadratic function. Robustness in the optical flow calculation is achieved by contrast content filtering. The algorithm successfully segments moving pedestrians from a moving vehicle with few false positive segments. Most false positive segments are due to poles and organic structures, such as trees. Such false positives are, however, easily rejected in a classification stage. The presented segmentation algorithm is intended to be used as a component in a detection/classification framework.