Satsuki Maeda, Bismark Kweku Asiedu Asante, and Hiroki Imamura, Soka University, Japan

Action recognition has many practical applications, but the task still faces significant challenges. A major challenge is the variation in human action poses across different viewpoints, which complicates determining the ideal pose for action recognition. To address these viewpoint-invariant issues, we propose a pose normalization approach combined with object-based action recognition to classify actions in videos. In this method, the normalized pose is compared with a reference pose to identify the action being performed. The objective of this research is to develop a three-dimensional (3D) objectassociated action recognition framework that leverages the stereo cameras ability to capture accurate distance information. This approach offers three main advantages: (1) action recognition that incorporates object context, (2) resolving occlusion problems, and (3) improving recognition accuracy through precise distance information. Experimental results show that our proposed approach achieves 70% classification accuracy across ten selected action categories, independent of viewpoint or camera angle.

Object Recognition, Behavior Recognition, AI, Stereo Camera, Active Detection