Comparative performance analysis of deep neural networks for human activity recognition
Abstract
Human activity recognition is a challenging computer vision problem, which involves identifying human activities using artificial intelligence methods, from the data collected from various sensors. In this paper, existing deep neural networks are studied and a custom deep architecture is proposed for human pose and activity recognition. The proposed architecture is mainly based on the approach of localizing human body joints in a 3D space on a single 2D image. The fundamental problem is the wide variations in appearance, caused by various camera angles which include, clothing, body shapes, and self-occlusion, and complications in background variations. The Human 3.6M and patient MoCap datasets are used as a benchmark and our proposed architecture shows competitive performance compared to the state-of-the-art methods.
Document Type
Conference Proceeding
Date of Publication
1-1-2023
School
School of Engineering
Copyright
subscription content
Publisher
IEEE
Comments
Burney, M. A., Waheed, A., Khan, S., Ghori, S., Shaikh, M. B., & Qureshi, R. (2023, November). Comparative performance analysis of deep neural networks for human activity recognition [Paper presentation]. 2023 18th International Conference on Emerging Technologies (ICET), Peshawar, Pakistan. https://doi.org/10.1109/ICET59753.2023.10374609