Multi-type Action Recognition Using Sparse Representation

• Main Authors: Zhe-Sheng Lu, 呂哲昇
• Other Authors: I-Cheng Chang
• Format: Others
• Published: 2015
• Online Access: http://ndltd.ncl.edu.tw/handle/12090139475230377751

碩士 === 國立東華大學 === 資訊工程學系 === 103 === Human activity recognition is an important topic in computer vision because it can be applied to various promising applications, such as video indexing and retrieval, human computer interaction and security. However, most of the existing studies only focus on the actions of a single group, and this study brings forward an automatic system which can recognize multi-group actions. The proposed system is based on two-level configuration: action group classification and action type recognition. The action group classification is to classify an input action into one of the action groups. Here, we adopt the motion vectors as the action features and Support Vector Machine (SVM) as the classifier. The action type recognition is to identify the type of the action within the group which is determined by action group classification. Since each action group has its own particular features, we utilize Pictorial Structures Model (PSM) to describe the variation of human body, and extracts suitable recognition features for the input action according to its action group. The recognition process utilizes a sparse representation-based method to learn the discriminative dictionary and recognizes the action. Two experiments are performed to show the efficiency of the proposed approach. In the first experiment, we evaluate the performance of the motion vectors feature and it performs well in the classification of action groups. In the second experiment, the performance of the two-level recognition system for multi-group actions is evaluated. Experimental results show that the system is able to recognize 30 different actions of three groups with natural backgrounds and the average recognition rate achieves 91.52%.