부산대학교 도서관

Human action recognition (HAR) in films has grown rapidly as an academic field in the past few decades. Robotics, HCI, intelligent video surveillance, and sports video analysis are just some of the real-world applications of action recognition. Despite extensive study, there are still many issues to be settled in this area. Multiple factors, such as the subject's position, speed, illumination, occlusion, viewpoint, and backdrop clutter, contribute to the challenge of the task. An efficient HAR system is able to account for these variants and rapidly identify the human action class. The primary steps in HAR systems are typically foreground segmentation, feature extraction, effective vector representation, and classification. This article proposes a novel method that employs Convolution Neural Network (CNN) to improve the HAR system's classification accuracy. The proposed activity representation and classification approach is evaluated by using public datasets from Weizmann, KTH, and the Ballet Movement. The examination of competing methods shows that our suggested approach provides higher recognition accuracy than existing methods. For the Weizmann dataset, the proposed technique offers 98 percent accuracy, and for the KTH dataset, it offers 95.6% accuracy.