Research On Human Behavior Recognition Algorithm Based On Skeleton Model

In recent years, the study of human behavior recognition has become a hot topic in the field of computer vision. Not only it can be applied to video surveillance or video retrieval system in public places. And with the development of intelligent robot, human behavior recognition has become an important technology of human computer interaction,which is a key bridge between human and robot interaction. Kinect sensor released by Microsoft extending the traditional two-dimensional feature description to the the 3D world including the depth information, promotes the study of human behavior recognition and makes it easier to realize the human computer interaction based on intelligent robot. The paper will study the human behavior recognition algorithm based on the human skeleton model obtained from the Kinect sensor.First of all, the noise in environment make the original wobble severely, so we propose to employ mean filter based on the human skeleton to smooth each joint’s movement. Secondly, in order to make the feature description based on the skeleton model be independent of the world coordinate system, we propose to do the normalization process, namely normaliz each frame of the skeletion to the local coordinate system.Then according to human mechanical theory,the human body model can approximate as a rigid body system link by revolutes, we decide to use angles and vectors,the simple mathematical concepts, to compute the characteristics of static postures and dynamic motion of human behavior. The identification model based on the Maximum Entropy Markov model is presented in the end. In addition, we also construct behavior datasets in indoor environment,respectctivly are drinking, stretching the body, mopping the floor, having a rest, working, standing.In the experiment, we make a contrast experiment based on the standard public dataset and the new behavior dataset. Accrocding to the complex environment in the real life,such as the light, noise and different camera views,we make a series of comparative experiments. By the experimental results, we get the the altoghrim we proposed have the advantages of high recognition rate, simple calculation, real-time performance, and robustness to illumination, noise, and different camera.