A block matching approach for movement estimation in a CMOS retina: principle and results

We present in this paper a silicon retina- a specific CMOS image sensor dedicated to movement estimation. This circuit has been fabricated in a standard 0.35/spl mu/m CMOS process. The pixel is composed with both sensitive elements- a photodiode is used to convert photons to electrons- and electronic components for computations. Its goal is to extract the optical flow from two acquired images. Movement estimation is a very often used task after image acquisition, which usually needs a sensor, a processing unit (DSP, FPGA) and external memories (M. Arias-Estrada and E. Rodriguez-Palacios, 2002). Here, the functionality is achieved in a single chip, by an electrical way, giving compactness and low power to this costing task. The movement estimation algorithm we use comes from (R. Zabih and J. Woodfill, 1994), its advantages are: robustness with real sequence, its real use in robotics, and its good electrical mapping. The circuit's architecture is divided in to two parts, corresponding to the two steps of the original algorithm.