Dynamic segmentation of traffic scenes

A new system for the detection and tracking of moving objects is presented. The system can be used to analyze image sequences taken either by a static camera or by a camera mounted on a moving vehicle. The system provides information about obstacles and relative motions for the assisted or autonomous navigation, and is based on a fast image motion estimation performed with a multi-scale/multiwindow algorithm able to compute a reliable optical flow even for large displacements between successive frames. When the image motion has been obtained, regions of motion discontinuity can be identified. This technique allows a fast but sufficiently precise detection and localization of moving objects. The detected objects are tracked by using "moving snakes", i.e. elastic contours initialized at the boundaries of the detected regions and driven in the successive frames by the first order approximation of the region motion, by a force generated by intensity edges and by an elastic force. A series of tests on different kinds of image sequences is presented, and the results obtained show that this approach provides a good segmentation of moving objects in different conditions.