Two systems for detecting the motion of a scene are described. For both, an image is projected directly onto an integrated circuit that contains photosensors and computing circuitry to extract the motion. The first system, which has been reported earlier, correlates the analog image with a digitized version of the image stored from the previous cycle. The chip reports the motion that corresponds to the maximum analog correlation value. This system represents an advance from previous designs but exhibits some shortcomings. A second completely analog design surpasses the first. The mathematical foundation is derived and the CMOS circuits used in the implementation are given. Test results and characterization of the working chips are reported. The new motion detector is not clocked and exhibits collective behavior. The use of local information extensively avoids the correspondence problem. The system can be thought of as a Hopfield neural net with one important extension—input driven synapses. The motion detector also meshes nicely with the existing computational vision work. Extensions to handle more complex motions are proposed. The suitability of the motion extraction algorithm as a biological vision model is explored.