7.1: Feature Detection as an Information-Reduction Problem

The information generated by sensors can be quite formidable. For example, a simple webcam generates 640x480 color pixels (red, green and blue) or 921600 Bytes around 30 times per second. A single-ray laser scanner still provides around 600 distance measurements 10 times per second. This is in contrast to the information that a robot actually requires. Consider for example the maze-solving competition “Ratslife” (Section 1.3) in which the robot’s camera can be used to recognize one of 48 different color patterns (Figure 1.3) that are distributed in the environment, or the presence or absence of a charger, essentially reducing hundreds of bytes of camera data to around 6 bit (2⁶ = 64 different values) content. The goal of most image processing algorithms is therefore to first reduce information content in a meaningful way and then extract relevant information. In chapter 6, we have seen convolution-based filters such as blurring, detecting edges, or binary operations such as thresholding. We are now interested in methods to extract higher-level features such as lines and techniques to extract them.