Research On Human Detection Based On Feature Learning In Depth Image

As a hot and difficult topic in the field of compute vision, human detection has be widely used in many practical applications, range from intelligent video surveillance, housekeeping service robots, assistant driving system, Machine navigation, intelligent transportation etc. Except the value of applications, human detection is a fundamental research for high-level computer vision technology such as human behavior recognition and human tracking.Depending on the type of image data, human detection is divided into RGB image and depth image. Currently, research of human detection based on the RGB image is mature since many proposed detection algorithms. However, affected by the change of illumination, shadow, occlusion and complex background, performance of RGB human detection is instability. As a new type data, depth image can not only save the information of object space position, but also have advantages such as personal privacy, illumination insensitive and less dimension. Thanks for the appearance of Kinect sensor, human detection based on depth image is becoming intense. Focus on the research of human detection based on depth image, this paper firstly take a brief overview at the basic idea of common feature extraction algorithms in details. Then based on the summary of human detection methods in the field of computer vision, we do some research on the following two aspects:1. In the speed of detection, we propose the concept of pre-detect cluster for the problem of huge search space which from the application of sliding window, and add it to human detection framework. In order to speed up the process of feature extraction, it positions the human body before detect and, Crop image set to be detected. 2. In the human descriptor on depth image, we introduce the method of deep learning into human detection based on depth image. Learn image depth characteristics and reconstruct depth information by sparse auto-encode (SAE).3. Present and implement Finally do experiments compare with SAE andAfter all, in the SZU Kinect People Datasets, we conducted a series of experimental, which compare the SAE feature with the typical algorithm of human feature extraction based on depth image, such as HOD, HDD, SLTP, RDS, and the improved SLTP algorithm. We also do comparative experiments between SVM and Softmax classifier.Experiment results show that the proposed framework and SAE deep learning for human detection in depth image are effective.