Abstract
The enhancement of sub-micronic technologies made CMOS image sensors an interesting alternative. Vision machines are traditionally based on architectures linked to an image sensor. The proposed approach in this work deals with the integration of photodetecting elements plus image processing circuits in a single chip. These smart sensors, also called retinas, enable to make the vision system both compact and optimised. The processing is indeed speeded up, and power consumption reduced, compared to classical solutions (sensor, then hardware or software processing). More precisely, this thesis deals with the integration of a motion estimation algorithm, using electronic functions for computations, instead of a strong iterative numeric computation. Afterf finishing a circuit for a correlation-based retina project, we first designed a standard sensor (an imager) to acquire knowledge in the image sensors field. Then, We designed a retina performing these computations in an electronic way. That motion estimation algorithm is based on a robust block matching technique, composed of a pixel coding step, then a search step of these codes in a potential destination window. That approach is new because it can treat textured scenes with changing luminosity, moreover we integrate a kind of algorithm that was only used in software or digital approaches (FPGA, DSP).