부산대학교 도서관

This paper proposes trajectory-based human action recognition using Center Symmetric-Local Binary Pattern (CSLBP) (structural descriptor). Histogram of Optical Flow (HOF) and Motion Boundary Histogram (MBH), computed along the trajectory of the action, have been taken into account as motion descriptors. Additionally, the trajectory shape descriptor is extracted. Harris corner is used to extract the key points. You Only Look Once (YOLO) has been employed to localize the human, for the KTH dataset, reducing the number of key points. Finally, the human actions are classified using K-means and Support Vector Machine (SVM) algorithms. The evaluation is performed on KTH and YouTube Action datasets, obtaining an accuracy of 91.6 percent and 90.4 percent, respectively.