Learning Human Pose And Action Similarity Metric Using Hierarchical Sparse Models

Video-based human body posture and action recognition is an important problem in computer vision and pattern recognition. Due to the movement of human action and the complexity of the data collection, body posture and action recognition process usually finds the best posture of the target body using a similarity measure between the input body posture and offline database a database searching algorithm. Thus, body posture and action recognition are critical to the accuracy of the similarity measure based database retrieval body posture and action recognition. The traditional pose similarity metric acquires human motion joint position posture using supervised learning or unsupervised learning algorithms. However, data from the human body posture movement joints consists of different levels of features, e.g., the head, hand and leg position reflects the people's station, waiting for a rough action, and the hand of the upper arm, lower arm, palm position the reaction of human fist, shook hands and other minor movements. The traditional pose similarity measure of body movement joints indiscriminately characterize expression, leading to the study of the similarity measure for different scales of operation attitude portrayed inaccurate. This thesis propose hierarchical sparse models based on unsupervised hierarchical sparse body posture and action similarity measure, consisting of the overall features and details of the various features minimize training samples stratified sparse on the training set represents residual, learning to be more accurate more robust body posture and movement similarity measure. The work of this thesis includes the following two parts.First, we propose a body posture similarity measure which is based on a hierarchical sparse representation model. We construct feature vector of the human posture using geometric relationships between the key points in body posture of single frame in the video. According to the difference of the importance for describing human posture using the body skeleton point, the feature vector is divided into rough features and detailed features, which construct the objective function building on the total residual of sparse representation in each training sample of the training set and get similar measure of body posture in the end. The experimental results on CMU Graphics Lab Motion Capture database shows that, our method produces higher accuracy and speed, compared with the geometric pose descriptor metric.Second, we propose an action recognition method which is based on Fourier spatial-temporal pyramid representation of human movement. According to 3D position of the body skeleton, we use the Fourier spatial-temporal pyramid model to get feature vector of human action, so that different actions have the same dimension for their feature vectors. Then we get the action similarity measure by the hierarchical model of the sparse representation. Experimental results on MSR-Action3 D database also demonstrate the effectiveness of our method.