Human Parts Recognition Based On A Single Depth Image

Human pose recognition is a highly significant research direction in the field of computer vision, and has been widely used in many industries, such as human-computer interaction, intelligent monitoring. Accurate identification of body parts in moving video can not only lay the foundation for human pose recognition, but also reduce the difficulty of human behavior analysis. In order to avoid the interference of illumination intensity, shadow, object texture and other external factors and reduce the impact of human body parts matching model’s error, the human parts recognition problem has been transferred into a per-pixel classification problem using the method of feature classification. In addition, to assemble pixels’ classified information, this paper achieves the estimation of human part joints using clustering algorithm and draws the 3D human skeleton as the final output of human parts recognition.The main research works of this paper are listed as follows:1. Because there is no public human part labeled sample database currently, in order to overcome lack of training data, this paper refers to several action sequences from CMU motion capture dataset, collect the sequences’ depth images by Kinect sensor and self-build human body parts labeled sample database using artificial mark. Furthermore, due to the human body model size may be affected by the human gender, weight, the difference of dress and other factors, the data source is sampled by several different body type tested people in real experiment scene.2. To improve the correct rate of body parts recognition, on the basis of depth comparison feature, introducing human body parts size factor, random forest algorithm is used to train classification model. The experiment shows that there is a higher human part recognition correct rate when use the improved classification model to recognize human parts.3. In order to predict human part’s joint and draw the 3D human skeleton, in this paper, firstly propose an improved Mean shift algorithm, which combine the body parts size factor to predict joints. Secondly, according to human physiological structure, extract the 3D map of human skeleton by means of connecting adjoining joints in straight line. Finally, the experiment proves that the improved Mean shift algorithm can increase the predicted correct rate of human part’s joint.