Motion segmentation using temporal block matching and LEGION

A motion segmentation method is proposed for an input sequence of random dot and binary images. The method is composed of two main stages, inspired by primate visual system. The first stage determines local velocity information at each location in every image frame using its two neighboring image frames. Measurements of a particular velocity at all locations form the corresponding velocity layer. The second stage performs segmentation based on the motion information in the velocity layers. Each velocity layer provides input to a LEGION (Locally Excitatory Globally Inhibitory Oscillator Networks), which is a 2D array of neural oscillators. When LEGION networks are simulated, the oscillators corresponding to the region of a uniform velocity oscillate in synchrony, whereas the regions with different velocities tend to attain different phases. Final output is displayed in the segmentation network. Results demonstrating the performance of our method on synthetic image sequences are provided and related to psychophysics.