Human Motion Capture Based On Monocular Vision

Human motion capture (HMC) is an important issue in the field of computer vision and machine learning, it has increasingly application value in the field of human-computer interaction, intelligent monitoring, virtual reality, animation, motion analysis.Though HMC based on computer vision has been researched for30years, but it’s still in the laboratory research stage, and can only capture a single person or a few people a simple action. The problem of automatic object extraction, occlusion, human modeling, pose estimation are still challenge, the robustness, accuracy, and processing speed are far from the practical application. To solve these problems, this paper focuses three main aspects:object extraction, human motion tracking, and pose estimation, the main work is as follows:First, we propose an adaptive object extraction method based on online K-means clustering Gaussian mixture background modeling, improve the traditional Gaussian mixture model initialization and updating algorithm. Our method can effectively deal with the backgrounds interference and extract object from complex scenes accurately and completely, and also has a strong adaptability for illumination change.Second, we propose a tracking algorithm based on compressive sensing with SURF, we take advantage of the stability of SURF and use a weighted classification ideological construct classifier. This method effectively avoids the complex issues of human modeling, achieve a fast and accurate tracking.Third, we propose a pose estimation method based on constrained tree PS model and location prior. First use a Grab-Cut to extract the foreground, which is used as inference region. After that, we learn the location prior of each part of the body to get the body’s appearance model. Then make an inference based message passing. Our method can solve the problem of huge search space of traditional model effectively, and improve the accuracy of pose estimation.Fourth, we construct a human motion capture system based on monocular video sequences. This system is a good platform to make a further study of human motion capture.