Human detection using geometrical pixel value structures

We propose a statistical method to detect human(s) in images by using geometrical structures that are common to the appearances of the target objects (human figures). Most appearance-based methods focus on pixel values directly, because the same classes of objects usually have similar pixel value distributions. However, this is not true for some particular objects. Humans are a good example. Human figures have a variety of different clothes, and their pixel values (color, brightness) can vary significantly from person to person. In this case, geometrical structures observed as pixel-value distances are essential for the successful recognition of objects. In this paper, we propose a method to describe and recognize the appearances of objects based on geometrical structures. The representation is based on a statistical analysis of Mahalanobis distances among parts of images. Using our method, objects having pixel value variety can be recognized using a small number of appearance models. Experimental results for human figures demonstrate the effectiveness of our method