Human Body etection Based on Template Matching Using Depth Image

Abstract

The task of detecting humans has been essential for many applications such as Human Computer Interaction (HCI), Human Robot Interaction (HRI), and surveillance monitoring system. In consequence, there have been a variety of human detection studies. Most conventional methods have used color images. Some methods using the color images employed background modeling or frame differencing to segment foreground and verify whether it is human or non-human by matching methods using shape information, modeling-based methods, or learning-based methods. However, these methods require efforts and time to segment the foreground. The other methods directly detected human by sliding window or patch without preprocessing aforementioned. However, it is hard for real-time applications because it needs long processing time. Besides, when background is complex or lighting is changed, these types of methods suffer from deteriorated performance of human detection. To solve such problems, recently methods based on depth image. These methods do not require to segment foreground from background using background modeling or frame differencing because it is possible to make blobs from pixels which have similar depth value. This paper proposes a human body detection scheme using ellipse template matching based on depth image. There are two steps for human body detection: human body candidate detection and human body verification. In human body candidate detection, the ROI is set to improve execution speed. To extract human body candidate regions, meaningless pixel blobs are removed using several pixel blob analysis technique as follow: 1) size of pixel blob 2) width of pixel blob 3) ratio of pixels consisting of pixel blob 4) position of pixel blob. To verify remaining human body candidate regions, human body verification is performed by matching ellipse template generated as considering shape and position of head. In the experiment, we compares the proposed method with conventional human detection based HOG (Histogram of Oriented Gradient) and method using 3D head model based depth image. The result of this experiment shows superior performance to complex background and rotation of head and body. Also, it shows the proposed method can run in real-time (20 fps)