Partially occluded human detection by boosting SVM

In this paper, a novel method to detect partially occluded humans in still images is proposed. An individual human is modeled as an assembly of natural body parts. Some part based SVM classifiers are trained first by using histogram of orientated gradient feature. Different from other boosting methods, region information is stored in each classifier. When detect human in crowed scene, according to the information of humans that have already been detected, the information of available regions could be obtained, when a new detection window is in process. In classifier sequence, the classifiers whose regions are available are selected for generating the final classifier. This method could achieve good performance on images and video sequences with several occlusions.