Coherent motion regions extend in time as well as space, enforcing consistency in detected objects over long time periods and making the algorithm robust to noisy or short point tracks. As a result of enforcing the constraint that selected coherent motion regions contain disjoint sets of tracks defined in a three-dimensional space including a time dimension. An algorithm operates directly on raw, unconditioned low-level feature point tracks, and minimizes a global measure of the coherent motion regions. At least one discrete moving object is identified in a time series of video images based on the trajectory similarity factors, which is a measure of a maximum distance between a pair of feature point tracks.