On SVM Decision-Based Human Activity Recognition Techniques for Single Camera Video

• Main Authors: Shih-min Liu, 劉詩敏
• Other Authors: Chin-shyurng Fahn
• Format: Others
• Language: en_US
• Published: 2011
• Online Access: http://ndltd.ncl.edu.tw/handle/8uc8jc

碩士 === 國立臺灣科技大學 === 資訊工程系 === 99 === In recent years, human activity recognition has become a popular research trend. Many scholars try hard in order to find an easy way for human activity recognition. But there exist many difficulties, one of the difficulties is human activities are continuity postures and each activity contains a lot of body moving parts; therefore, those activities involve a high dimensionality and complexity. Another one is that researchers need to deal with the spatial and temporal information of frames in each video sequence. The other one is when more than two people appeared in one video, researchers need to assign an ID for the actors in the video so as to correctly recognize activities for each person. In this thesis, we proposed human activity recognition techniques based on the decision method. After the preprocessing steps, we obtain the minimal bounding box of body from the video. Once we get parameters of the bounding box we put them into equations after that we can get some feature vectors. Then we use Hu moments, body orientation angle and the feature vectors to construct a Support Vector Machines (SVM) decision tree. Each SVM on the decision tree is trained and tested separately to achieve its best classification performance by choosing proper features and parameters. In the training step, we construct our training data which is manually determined by observers. Our human activity recognition technique can detect non erect and erect continuous postures. We also propose a simple ID assignment method to detect more than one people in the same time under our experimental restrictions. In the experiment we use video sequences with resolution of 320ⅹ240. The recording equipment with such resolution usually is cheaper than high resolution equipment, so it is good for smart home environment surveillance and other applications. Our system could recognize six kinds of human activities, including “walk”, “run”, “squat”, “sit”, “jump” and “lie”. According to the experimental results, the average recognition rate for one person is 97.1%. As for multi-person activity recognition with our experimental restrictions, the average recognition rate can reach 95.9% with at most 5 frames delay periods.