Research On Human Behavior Recognition Based On RGB-D Images

With population aging father advances, the phenomenon of elderly people living alone is increasingly intensifying. The home service robot with the main task of protecting and helping older adults is becoming an important research field of artificial intelligence. How to understand of human behavior is the key of the home service robot. Human behavior recognition as a basic research catch more people’s attention and more and more people begin to study human behavior recognition. Therefore, in order to accurately and quickly to recognize the human behavior, and overcome the environmental impact of existing methods, we try to use the depth images to recognize the human behavior. With analysis of the existing approaches, we proposed take advantage of the depth information into human behavior recognition. Combining the depth information with human behavior recognition, the paper proposes a recognition method based on RGB-D information. And focus on the following aspects.Firstly, the depth image obtained by kinect is analysed. The paper analyzes the forming reason on hole in the depth images. And we proposed fixes algorithm of the hole, by preprocess the image achieved hole filling and noise.Secondly, analysis the edge detection algorithm based on the RGB-D image. Primarily we studied several classical edge detection algorithms. And determine the edge detection algorithm in the process of the human behavior recognition algorithm. Since there are some holes and noise in the depth image acquired by Kinect. And the hole and noise caused the edge information extraction is not ideal,we proposed an algorithm to repair the edge of the depth image. Then based on the ‘ ? ’template to positioning the head in the depth image. We proposed an algorithm of template matching, using MAD to positioning the head based on depth image.Thirdly, this paper proposes a method based on RGB-D information of region growth. Using the proposed region growth algorithm based on depth image to extract the human region. And adopt the HOG feature to extract the characteristic of the depth image. Using the characteristic to train classifier of GRNN, and realized the goal of the human behavior recognition.Finally, experimental results show that the proposed method is feasible and effective, and analyze the experimental results.