| Visual motion information is a key primitive for dynamic scene understanding. As a first step, the image bidimensional velocity vector field, known as optical flow, has to be computed. Based on the optical flow, it is possible to compute visual feedback, collision avoidance, time-to-contact computation, depth and structure estimation and 3-D navigation tasks, which are particularly useful for mobile robotics and autonomous vehicle applications. In those contexts, real-time performance is highly desirable, but unfortunately, the estimation of dense motion field is computational intensive. Real-time systems, based on the traditional camera-computer approach does not lead to practical solutions in terms of size, power consumption and cost.; Compact and low-power vision systems can be developed by carrying out low level visual computations at the sensor level, before delivering information to a host system. Integrating computation capabilities and sensing in a single module is considered the most promising technology to implement compact and cost-effective vision machines. VLSI technologies, like standard digital integrated circuits CMOS (Complementary Metal Oxide Semiconductor), allow the implementation of photo-transduction, analog, and digital signal processing modules in the same substrate. There are inherent difficulties on the design of a motion sensing architecture due to the complexity of the algorithm and the restriction of silicon space per sensing pixel.; In this thesis, a focal plane architecture for direct computation of visual motion was developed. The architecture is based on theoretical and practical considerations and it improves previous approaches. The proposed architecture consists of a bidimensional array of intelligent pixels working in parallel. Each pixel has a photo-transduction element, analog signal processing and a digital communication channel. Individual pixels have self-communicating capabilities to signal their coordinates when motion is detected. The pixel array detects moving features in the image, and communicates them through an asynchronous digital communication channel with a time-stamp protocol to a companion module. 2-D velocities are then computed as a second step in the module. Visual Motion computation is thus simplified by an image sensor with motion detection features and a velocity computation module. The approach permits the implementation of high resolution sensors with current microelectronic technologies.; A VLSI circuit in CMOS technology was designed, simulated, fabricated, tested and the whole architecture was validated. The sensor functionality is demonstrated in a prototype camera controlled by a high performance microcontroller, and connected to a host computer through a serial link. The host computer receives the motion information for higher level analysis. The concept can be extended to multifunctional vision chips and low-power, compact, and computational efficient smart cameras for robot vision and autonomous vehicles. |
