Hands and Face Detection using Skin Blob Extraction
We use parameterized Mixture of Gaussian (MoG) distribution to represent Generic Skin color model for humans. We used normalized RGB to reduce the shadowing effect and illumination variations. We use Viola-Jones face detector to extract face bounding for subjects in the first 20 frames of the video. We refine (personalize) the Generic Skin Model from the extracted face regions as most of the pixels in the detected rectangle are skin pixels, and the probability of a non-skin pixel to have value similar to skin pixel is very low. We use first 20 frames of the subject's face to learn the new skin model is specific to the person's skin color. Based on the new skin model we detect all skin regions, i.e. face and hands in rest of sequence. The process runs in real time at 25 frames per second.
Shoulders detected by finding lines around Face with orientation in a range. To detect shoulders we look in a bounded rectangles on each side of Face detected using Viola-Jones face detector. We use Progressive Probabilistic Hough Transform (PPHT) for Line detection. PPHT is computationally efficient and generates fewer false positives. PPHT generates Line Segments instead of lines (Standard Hough Transform), by randomly sampling pixels from the image and adding them to histogram bins. PPHT allows line selection based on minimum threshold length. We detect Lines using Progressive Probabilistic Hough Transform and prune off lines detected with unreasonable orientations. Finally we merge the multiple segments.
Palm Detection Click on the images to view the video
In order to detect palm regions in the arms, we extract the edges inside the
detected skin regions, using the Canny method. We estimate the edge densities;
in cases of individuals with short-sleeve clothes, we segment the hands from the
arms based on the hands’ increased edge densities (compared to the arms’ edge
densities). In this way, we extract the blobs of the palm regions and
not the entire arms.
Shoulder Shrug Detection Click on the images to view the video
Other Results Click on the images to view the video
Description about the project [Link]
Dynamically Adaptive Tracking of Gestures and Facial Expressions, D. Metaxas, G. Tsechpenakis, Z. Li, Y. Huang, and A. Kanaujia, ICCS 2006