Pulse coupled neural network for motion detection
Bo Yu, Liming Zhang · 2004
This paper presents a pulse coupled neural network that segments moving objects from background. The model is composed of Eckhorn's spike neurons arranged in two parts. Part I is a two layer network that performs local features matching. Part II is one layer of local connected neurons inhibiting false matching. The visual input is encoded in pulse sequence, whereas the motion direction and distance are built into synapse delay. The object contours within two consecutive frames of input video match each other through neurons acting as coincidence detectors. If an object is moving, its contours in consecutive frames will be fully matched; the neurons included within these contours will fire periodically. The contours of the still object will be inhibited by the false matching removing mechanism in part II. Finally, only the moving object(s) emerges from the background.