Advanced Human Detection Using Fused Information of Depth and Intensity Images

Abstract

Human detection systems have been applied to many applications such as intelligent vehicles and surveillance cameras with increasing demands on safety and security. The scope of previous works has been confined usually in color (or intensity) images. In this chapter, we present a complete human detection system using the information on both depth and intensity images. First, we apply a segmentation algorithm to a depth image. Then we merge the segmented regions and generate Region-Of-Interests (ROIs) which may contain a human, considering experimentally determined horizontal overlap and aspect ratio, respectively. Second, we use a newly proposed feature descriptor, Fused Histogram of Oriented Gradients (FHOG), to extract feature vectors from the ROIs applied in both depth and intensity images. Finally, we check the presence of humans in the ROIs with linear SVM. Following the basic principles of Histogram of Oriented Gradients (HOG), we develop this FHOG descriptor to utilize both gradient magnitudes of depth and intensity images. With our datasets obtained from Microsoft Kinect sensor, the FHOG descriptor and overall system achieve a miss rate of 1.44 % at 10−4 FPPW and of 10.10 % at 1 FPPI, respectively. The computing time of proposed system is also significantly reduced. Experimental results show our system is able to detect humans accurately and fast.