People detection in still images based on a skin filter and body part evidence

This paper presents an approach for determining the presence of people in a still image. The approach starts with the segmentation of skin areas in the images using the YCbCr color space and a set of experimentally defined thresholds. The detected skin areas are then grouped into body part candidate regions. Regions that are too small (typically smaller than 5x5) are discarded. A simple heuristic rule, based on the proportions and pixel density of these regions, is then applied to map these regions into body part evidence. The application of this approach to a database of still photographs resulted in the correct classification of 77.2% of the images.