Human detection in depth image sequences based on CoHOD features

The spatial clues in the new medium of depth image sequences offer the potential opportunities to solve difficult problems in human detection, such as varying illumination, approximate appearance and occlusions. Therefore, object detection based on depth image sequences has attracted much attention in recent years. The key problem of human detection based on depth images is the efficient extraction of discriminating features. Some existing methods have shown efficient performance in human detection in a given scene, such as torque characteristics, geodesic features and histogram features. However, these methods still have some disadvantages in robustness, accuracy and computational complexity because a depth image always contains noise and the human body includes detailed structures. In this paper, we propose a human feature representation method, called CoHOD (Co-occurrence Histograms of Oriented Depths), which combines the advantages of HOG (Histograms of Oriented Gradients) and co-occurrence matrix methods. In the proposed framework, a co-occurrence matrix is used to represent the human body in a depth image, through which the human body ultimately is efficiently and accurately detected. Experiments are conducted to demonstrate the effectiveness of our proposed method. © 2014 Binary Information Press