Human Hair Segmentation and Length Detection for Human Appearance Model

This paper presents a new approach for human hair length detection. In contrast to others, the proposed method is able to segment hairs from different views of human heads even with low resolution. Faces are not necessary to be visible in the images, as no face detection is needed in our method. Firstly, it conducts background subtraction to detect foreground objects and then detects human heads with a trained human head detector. Next, histogram analysis is carried out on the detected head region to segment hair region with K-mean clustering. Finally, hair length is determined by performing line scanning on the segmented hair region. The main advantages of the proposed method are able to handle the cases of (1) view variation and (2) low resolution or small human body appearance in the image. It is specially designed for surveillance system. The detected human hair length can be used as an invariant feature to be embedded into a human appearance model which can be employed for human detection, indexing and searching in multi cameras network system. This method aims at the fast processing speed and is able to achieve 7ms for hair length identification in a head patch detected. The experiments show the improvements in terms of speed and accuracy.