Human Action Recognition in Videos Using Inception-v4 Deep Learning Model
Divya Rani R, C. J. Prabhakar · 2024
The goal of video-based Human Action Recognition systems is to automatically analyses and recognize the actions performed by human being in videos. HAR is extensively applied in numerous applications such as, intelligent surveillance, video storage and retrieval, robotics and healthcare. The most challenges associated with action recognition are cluttered backgrounds, illumination variations, occlusion, intra and inter-class similarities. Recently, Deep learning-based techniques have shown highest accuracy for recognizing human actions. Hence, this paper is initiated with review of many deep learning-based techniques proposed for recognizing human actions in videos. More emphasis is given on CNN based techniques. In this paper, we proposed a deep learning-based method for recognizing human actions present in videos by employing pretrained Inception-v4 model. The proposed method is assessed by using three benchmark action datasets: KTH, UCF-101, and HMDB-51. The results obtained from experimentations with these datasets for the proposed method shows that, the proposed method exhibited outstanding performance compared to other state-of-the-art deep learning-based HAR techniques. The proposed method produced recognition accuracy of 97.83%, 95.48%, and 84.43% for KTH, UCF-101, and HMDB-51 datasets respectively. It is evident from the experimental results obtained from the proposed method is that the proposed method performed well in handling all kinds of challenges present in action videos.