Feature extraction from a depth map for human detection

Abstract

Human detection is challenging and important task for computer vision-based researchers. Histogram of Oriented Gradients (HOG) descriptors show significantly outperforms the existing feature sets for a human detection. Machine learning algorithms based on HOG have been widely used for classification, human detection and recognition. However this method has the amount of type I and II errors. The human detection using a visible light image was often affected by the change of illumination conditions, the pose and coordinates change, the interference of complex backgrounds, the camera sensor noise, the shadow of detecting objects and moving objects of self-occlusions or mutual-occlusions. The amount of these errors can be decreased by using the object distance information. In this paper we propose a human detection method based on HOG features. The method uses a depth map instead of visible light image. We have additional information about the scene through the comprehensive distance information analysis. During the experiment, we used maps of depth received from the Kinect v2 visual sensor. During the first step in our detection experiment we processed the whole depth map frame with the preprocessing noise redundant filters. Then on the second step we sliced the depth by the layers which contain definite distance points. After that step we made segmentation on the every layer and send segment to the pre-trained HOG-classifier. The experimental results show that the new proposed method of HOG on the depth map provides high precision and recall. It gives opportunities to train more sensitive classifiers, which can provide the higher recall values.