Implementation of Depth-HOG based Human Upper Body Detection On A Mini PC Using A Low Cost Stereo Camera

In this paper, we propose a human upper body detection using a depth image that is implemented on a mini PC using a low cost stereo camera. The camera, named Minoru 3D webcam, produces depth image from its two parallel cameras. A pyramid-based region of interest (RoI) is applied on the depth-image frames to scan the possibility of human upper body existence. Simultaneously, a histogram of oriented gradient (HOG) method is performed in the ROI to extract the HOG feature. The results of the HOG feature extraction are then classified using linear support vector machine (SVM). Our system has been experimentally tested publicly in the campus environment. The detection speed obtained from a computer is 8.53 fps and the Mini PC is 4.26 fps under non-threaded programming. The result of object detection using HOG and SVM Classification method on static image reaches an average accuracy of 78.71%. Testing the system for real implementation has average error accuracy 4.60%. The results of detection for two human objects can achieve an accuracy 71.00%. Moving image testing has an average error of 4.60%.