Motion stream analysis based on perceptual feature partitioning and grouping

We present a perceptual organization based on method for motion stream analysis. The computation model was developed based upon a perception principle: visual feature partitioning and grouping. In the method, perceptual edge features are extracted and classified into generic edge tokens (GETs) using edge tracking and partitioning on the fly. GETs are perceptually distinctive features of lines and curve segments. Various structures and patterns of GETs can be grouped in terms of the rules of perceptual organization laws. GETs are descriptive and therefore can be manipulated qualitatively. For each consecutive image pair, motion GETs (MGETs) are segmented by directly subtracting the GETs extracted in the first image from the same locations in the second image, in that no explicit GET pair matching is needed. The MGETs are then grouped into clusters based on selected rules and domain knowledge of the objects. The motion clusters are evaluated using the measure of motion persistence (over multi-frames) for eliminating unstable data, i.e. noises. Two result demonstrations include road mark following and vehicle tracking.