Research On Human Behavior Recognition Algorithms Based On Depth Information

Human behavior recognition has important applications in video surveillance andhuman-computer interaction, and it is also an important and challenging issue. Mosttraditional behavior recognition focused on RGB image sequences, but RGB imagesobtained from visible light camera were subject to interference of light, shadow andenvironmental changes et al., which affected the performance of behavior recognitionalgorithms. With the advancement of Kincet sensor, the cost to obtain the depthimages rapid declines and it will be much easier for us to capture depth image, whichcan provide3rd dimension information to help us represent actions more accurately.After that, many scholars paid much more attentions into researching and applyingdepth images. Unlike the visible RGB images, depth image pixel values only meansthe spatial position of the object, thus, the influences of light, shadow, color, climatechange and other factors on the depth images are small, thus, according to the depthinformation can solve the problem of segmenting body from RGB image to someextent. Therefore, in this paper, we study the problem of human behavior recognitionbased on depth information and a combination of depth information and RGB images,where our researches focus on describing motion change process and behavior featureextraction. The main results are as follows:1. In order to capture human motion change process better, we firstly defines anovel motion history image based on depth information, namely, the depth differencemotion history image, and then propose a new method for RGB figure foregroundsegmentation which adopts depth map multiply the gray-scaled RGB image to get thegeneral area of the human body, after that, human motion change process from RGBvideo sequences are captured, namely, the depth limit RGB difference motion historyimage, which can represent human motion change process much better thantraditional motion history image.2. Since an object can obtain different information from different perspectives,thus they also play an important and complementary role in behavior recognition,which can help to improve the accuracy of behavior recognition. Therefore wepropose an effective and efficient approach to depict human behavior to improvehuman behavior recognition accuracy, namely, axonometric projection (frontaxonometric projection, top axonometric projection and left axonometric projection).3. So far, human action recognition based on the depth information has madesome achievements, but it still does not have robust descriptors, and their accuraciesstill cannot satisfactory, therefore, we proposes a hierarchical block mean feature,which can better capture the spatial distribution of human movement, at the same time, we also successfully introduces GIST and PHOG descriptors, applied to scenesegmentation and classification, as our behavior descriptor. Experiments results showthat our descriptors have strong robustness, distinction and stability, whoseperformance is much better than that of the state of the art algorithms.4. Although the depth images have a variety of advantages, but the visibleimages also have many unique and complementary advantages. Therefore, differentdepth features and RGB feature for behavior recognition are fused in this paper,which can further improve the accuracy of behavior recognition. Experiments showthat the performance of fusion descriptors is much higher than only using depthinformation or RGB image, and our best performance on DHA dataset reaches95.2%.