Because we have restricted our attention to wooden office doors in the Gates building environment, we assume that all the doors we consider have, for the most part, similar hue and saturation values. We assume that in circumstances where this is not quite the case (example, under adverse lighting conditions), a substantial enough portion of the door has the right color characteristics so that we can recover the doors shape. Similarly, we assume that any occlusion that occurs is not substantial enough to render shape-recovery impossible. We also assume that other objects in the environment having similar color characteristics are not door-shaped.
Assuming we have accounted for lighting, we need to recover the shapes of the wood-colored objects in a scene. As we will show, noise, quantization, and distortion, ubiquitous headaches to machine vision in general, make this process difficult. Then, having determined an object's shape in the image, we must deal with the geometrical problem of correcting for projection and perspective transformations before being able to decide if it is a door or not.
Finally, we felt it important to try and develop solutions that could easily scale to problems with other kinds of doors, other features, and other environments.
We encourage the reader to keep these issues in mind while examining our proposed algorithm. The reader should notice, though, that we do not deal with all of the above-mentioned issues. For example, we do not correct the image for distortion, nor do we completely check for geometrical constraints (we check only for parallel lines that are perpendicular to image horizontal axis).
Next: Color filtering