Action Recognition Based On Kinectdepth Maps

Human behavior recognition is a widely research topic in computer vision, pattern recognition, image processing and artificial intelligence field, because of its wide application in the field of intelligent security and human-computer interaction. The traditional recognition methods mainly focus on two-dimensional image or color video sequence. There are many classical image processing methods are proposed. However, due to the defects of 2D image data, these algorithms are vulnerable to changes in illumination, occlusion and other environmental factors, and the recognition rate is not too good. With the development of computer vision technology and hardware, people found the depth images obtained by K inect devices are not affected by illumination changes, occlusions and changes of view etc. There fore, more and more researchers focus on the study of the behavior recognition algorithm based on depth image.Extracting the key information which can represent the human behavior from the video sequence or image is an important step of behavior recognition. Based on a lot of research on the behavior recognition of depth image, this paper provides a new method for the research of behavior recognition, and the research contents are as follows:First, based on the depth image motion history image(Motion History Image MHI) and static historical figure(Static History Image, SHI) action representation method were introduced. We optimize the motion history image and static historical image makes it able to retain more critical motion information. After feature representation, the LBP feature is extracted from the feature template. And the SVM(Support Vector Machine) classifier is used to classify the feature vectors. The effectiveness of our proposed method is demonstrated by simulation experiments on a public database.Second, we propose a sub- motion history image based on adaptive partition of motion energy. According to the accumulation of energy, the algorithm can be used to divide the whole behavior into several sub-actions, and get the motion history image of each sub-action. Then, the HOG features descriptors and LBP features descriptors are extracted respectively from the feature template, and the SVM classifier is used to classify the feature vectors. The comparison of the experimental results shows t hat the proposed method can better capture the key motion information of the representation, especially for the high motion similarity.