| Artificial devices that are active and participate in the real world are dependent on their sensor systems to provide a snapshot of the environment around them. This vital connection between external reality and internal representation can determine the entire character of how a machine interacts with its surroundings. For systems whose only purpose is sensing, this is the beginning and end of their abilities. Even for many varieties of robotic applications, it can be the limiting factor on performance. Visualizing the environment in the three spatial dimensions is commonly the most desired capability.; In all but the most specialized applications, often sensor processing methods attempt to convert specific sensor signals into the most generic digital representation possible, followed by digital processing at some location likely distant from the sensor itself. Digital processing allows some flexibility in processing specifications, but comes with a price. System resources can be wasted by trying to most details of the raw data in digital form for long-distance communication. The entire process involved for "packaging" sensor data for digital computation risks losing essential qualities available at the sensor. It also incurs penalties in performance and power efficiency.; I demonstrate in this thesis two sensor processing systems following a contrasting philosophy from such traditional methods. Sensor processing is carried out as near to the physical sensors as possible, capturing essential features immediately. By also concentrating on specific representations of data designed to be most useful, efficiency is the result. A visual imager and sonar array system are shown to and yet use little power. They carry out their functions more in a superior way for certain applications than presently existing alternatives. In some cases, they enable new applications not previously possible because of extremely optimized performance. |
