Research Of Gesture Recognition And Human Body Pose Tracking Algorithm Based On Computer Vision

ABSTRACT:Human motion analysis is an important research domain in the computer vision field. It has a broad prospect of applications. But because of the complexity of the human body posture and the limitations of visual theory, most of the computer visual analyses with human activities centered are still in research stage, the actual application occasions are not too much. In this paper gesture recognition and human body pose tracking algorithm of human motion analysis were studied.A fast and practical static gesture recognition algorithm was proposed in this paper. The backgrounds regions with similar color to the skin were removed by combining the motion information and the skin color. By this way gesture images are split out accurately in HSV color space. The problem of the rotation, scaling and translation in gestures recognition was solved by extracting the Hu invariable moments of the gesture images as the feature vectors. Static gesture recognition was completed by calculating the distance of the Hu moments feature between the input gesture image and the template gesture image. Experimental results demonstrate that this algorithm is real-time with the average recognition rate of88.2%.In the research of dynamic gesture tracking, Camshift algorithm was used considering the skin color characteristic. In this paper, the color parameters after skin color segmentation will be the input of back projection step of the Camshift algorithm. By this way the Camshift algorithm was improved to track the dynamic gesture automatically in real time and the algorithm was effective for the changed gestures.In the research of human body pose tracking algorithm, annealing thought was used in the particle filter algorithm to solve the problem of large number of particles. Human body3D cylindrical model was established using the data from the public human motion database HumanEva. The edge and silhouette image features were calculated to construct the likelihood function. And the angle range of the human body joint movement and the self-intersection detection were joined. Suitable annealing particle number and the layer number of annealing were selected through the contrast experiment to balance the tracking efficiency and tracking accuracy.Dimension reduction algorithm was studied in this paper to solve the problem of dimension disaster in human body pose tracking. Principal Component Analysis (PCA) method and Gaussian Process Latent Variable Model (GPLVM) method were studied in this paper. Stochastic gradient descent method was used to replace the original optimization method in the GPLVM. Experiment on public human motion database CMU shows that the learning results and learning efficiency of the improved GPLVM algorithm are better to the traditional GPLVM method and PCA method, and the data in the latent space has the effect of clustering.