Research On Marker-Less Human Motion Capture

Marker-less human motion capture is a highly active research topic in computer vision.The research topic has a wide spectrum of potential applications, such as intelligent video surveillance, human-computer interaction, animation making, athletic performance analysis, content-based video retrieval, etc. In intelligent video surveillance, marker-less human motion capture is a challenging task, which not only needs to be accurate, robust, real-time and without human intervention,but also needs to overcome the difficulties of cluttered background, occlusion and self-occlusion, ambiguities from3D-2D projection,etc.Aimed at the application of intelligent video surveillance, vision-based marker-less human motion capture is researched and discussed in this dissertation.This dissertation focuses on three important issues in marker-less human motion capture, that are human detection based on monocular vision, human action recognition based on monocular vision and human pose estimation based on binocular vision, proposing a series of methods and algorithms. The main contributions of this thesis are summarized as follows:1) In the human detection based on monocular vision, a fast human detection method in complex scenes is proposed. Aimed at the complex scenes with more serious occlusion, this method effectively eliminates camera shake, noise and complex background disturbance through image match, background subtraction with adaptive threshold and improved morphology operation; and effectively overcomes the occlusion problem through object discrimination based on head feature.For improving the applicability and accuracy of human detection, a human detection method based on adaptive background model and body parts is proposed. This method extracts foreground motion objects through adaptive gaussian mixture background model. This method removes the false object and completes human detection through two-step object discrimination algorithm, which includes rough body localization and accurate body localization based on head-shoulder parts feature in each connected component detected. Experimental results show that the two methods of human detection we proposed can quickly realize real-time human detection with high accuracy and strong robustmess.2) In the human action recognition based on monocular vision, an improved human action recognition method based on motion features and spatio-temporal features is proposed. This method aims to improve the accuracy of action recognition. Optical flow features and3D Histograms of Oriented Gradients (HOG3D) as motion features and spatio-temporal features respectively, we extract the local spatio-temporal maximum value of optical flow in motion human region of the video and HOG3D descriptor of spatio-temporal interest points in the video as feature vectors. Based on bag of words model, these two types of feature vectors are used to creat visual vocabulary and produce histograms of visual word occurrences respectively.When creating visual vocabulary of local spatio-temporal maximum value of optical flow features, considering the relevance of human action and spatial position, the improved human region sub-blocks method based on human structure is used to creat visual vocabulary.Two types of histograms of visual word occurrences we receive is trained and classified by the theme model of probabilistic latent semantic analysis.The final result of action recognition is obtained through a weighted concatenation of two features classification results.This method makes up the deficiency of two types of features. Experiments on both KTH action datasets and WEIZMANN action datasets show that our proposed human action recognition method is able to recognize human actions with high accuracy.3) In the human pose estimation based on binocular vision, a set of human pose estimation method is proposed. In order to alleviate the impact of the occlusion problem and ambiguities problem from3D-2D projection, human spatial depth information is introduced to improve the accuracy of human pose estimation method. After using two-stages method based model plane accomplishing camera calibration of binocular stereo-vision system, an algorithm of stereo matching and depth information acquisition based on human body feature points is proposed for acquiring human depth information.This algorithm detects human body feature points through Haar-like features combined with star model, then uses an improved stereo matching algorithm based on feature points and region matching to do feature block matching, finally acquires the accurate3D spatial coordinate of human feature points through camera calibration results.This algorithm has fast response and high accuracy, which performs twice positioning for each pair of feature matching points based on local matching algorithms.In order to describe human pose more accurately, a pose estimation algorithm based on3D spatial information of human feature points is proposed. Position-based spatial relationships of human key feature points as features, this pose estimation algorithm uses example-based method to estimate human specific pose. This pose estimation algorithm can more accurately estimate various pose for the introduction of depth information. Experimental results show that the set of human pose estimation method based on binocular vision we proposed can accurately acquire the3D human pose data to accomplish accurate human pose estimation, which effectively reduces the impact of the occlusion problem and ambiguities problem, improves the robustness of human pose estimation without human intervention and initialization.