Research On Human Posture Detection And Motion Recognition Technology Based On Kinect

Human posture and action recognition has been a topic of active research direction in the field of computer vision, it involves image processing, pattern recognition, artificial intelligence and the cross-link of various disciplines. It has been widely used in human-computer interaction, intelligent video surveillance, health care, virtual reality and motion analysis, and many other fields. However, there are many basic problems have not been resolved in the traditional 2D vision methods. With the development of image acquisition technology, especially the emergence of depth sensing cameras named Kinect with affordable price. It provides a new opportunity for gesture recognition researchers. Currently, depth image is mainly used for 3D scene reconstruction, while it is still primary stage in gesture recognition field. Although Kinect SDK provides skeletal tracking API, but it does not provide gesture recognition API. The main problem is the ever-changing of the human posture, so it is difficult to build a general recognition model.This paper use Kinect to obtain 20 human skeleton joints with the help of released SDK, mainly focus on the recognition technology of human typical posture and interactive action. Firstly, a joint-distance and joint-angle feature model is proposed based on the analysis of skeleton data. The model is not complex, easy computing and compositionality. It is verified by experiment that the model can keep scaling invariance so that entirely meet the needs of the research. Secondly, this paper investigates daily-life typical static posture through questionnaires to build posture library. A program is designed and developed for collecting posture sample. Then the posture recognition algorithm is designed based on KNN, and the experiment results show the proposed method is feasible. Finally, in the aspect of action recognition, the method of the action representation is proposed on the Kinect platform. Action recognition is spatiotemporal complexity compared to posture. This paper solves this problem combined with dynamic warping thought. The action recognition algorithm based on DTW is proposed and the action recognition experiment system is implemented. Six kinds of hand gestures are defined based on the interaction. The experiment results show that the proposed method can achieve a high recognition rate in the homemade action library. The results prove that proposed method is effective, robust and able to meet the needs of interactive tasks. Meanwhile, the system is valuable for training and recognizing action.