Dedicated frontends for embedded vision systems

Industrial image acquisition and processing tasks often make high demands on vision systems. The frontend implementation of CMOS image sensors with integrated preprocessing presents a feasible alternative to conventional CCD systems. To cope with multiple constraints such as time-to-market, power-consumption, size, costs as well as the typically high system complexity, an appropriate design methodology is required. In prior work, a holistic design methodology was developed with focus on embedded vision systems. In our investigations of representative vision applications several relevant processing tasks have been identified that are excellent candidates for implementation in the vision system´s frontend. Their mixed-signal hardware realization in the context of the holistic modeling approach is the focus of this paper. Aspects such as data reduction by integrated image preprocessing algorithms and feature computation, local adaptation, and electronically controllable foveated sensor planes as well as high-dynamic, linear response pixel cells are presented.