23 Robust and Efficient Robot Vision Through Sampling

Vision is an extremely important sense for both humans and robots, providing detailed information about the environment. A robust vision system should be able to detect objects reliably and present an accurate representation of the world to higher-level processes, not only under ideal conditions but also under changing lighting intensity and colour balance; when fully or partially shadowed; with specular and other reflections; with uniform or nonuniform backgrounds of varying colour; when blurred or distorted by the object’s or agent’s motion; in spite of chromatic and geometric camera distortions; when partially occluded; and under many other uncommon and unpredictable conditions. Visual processing must also be extremely efficient, allowing a resource-limited agent to respond quickly to a changing environment. Each camera frame must be processed in a small, usually fixed, amount of time. Algorithmic complexity is therefore constrained, introducing a trade-off between processing time and the quality of information gained. Within the domain of the RoboCup four-legged league, previous vision systems have relied heavily on the colour of objects since the ideal colour of most important objects is specified in the league rules: a green field with white lines, an orange ball, yellow and blue goals, and pink, blue and yellow navigational beacons. However, there is considerable scope for interpretation and variation allowed by the environmental specification, particularly with regard to lighting intensity and uniformity. Agents must be capable of performing under varying conditions, albeit with time allowed for detailed calibration procedures. Typical systems group the continuous space of colours returned by the camera into a small set of discrete, symbolic, colours. They then attempt to form objects by grouping neighbouring similarly-classified pixels (Bruce et. al., 2000; TecRams, 2004). In implementation this usually results in a look-up table or decision tree that quickly maps the detected pixel value to a symbolic colour. Typical approaches to the generation of this table involve a supervised machine learning algorithm, where a human expert provides classification examples to a computer program, which generalises these to form the complete segmentation (Pham, 2004; Röfer et. al., 2004; Veloso et. al., 2004; Brusey & Padgham, 1999; Xu, 2004; Chen et. al., 2003).