Multi-view Human Head Detection in Static Images

Detecting humans in images is an important task in image or video processing field. This paper proposes an illumination and poses invariant method for human head detection. The principle stands on the evidence that human head has similar shape, which is composed of several edge segments at certain orientations on head contour. Firstly, head contour are labeled manually on the head sample images, then contour image is extracted and transformed into gradient space, finally gradient image is decomposed into several scalar magnitude images each with the same quantized phase. Standard boosting algorithm is utilized to search for weak classifiers. To our knowledge, we built the first multi-view head detector in the world with gradient information only. Experiment testifies good performance of proposed method.