Human Action Recognition using Attention EfficientNet
Bhumika Karsh, Rabul Hussain Laskar, Ram Kumar Karsh · 2023
The integration of Human Activity Recognition (HAR) within automated residences and surveillance systems has significantly elevated the significance of this field for advanced investigation. HAR entails the deployment of artificial intelligence (AI)-driven algorithms to autonomously discern human actions. Given the remarkable potency of deep learning (DL) methodologies in deciphering intricate tasks, their supremacy over conventional machine learning (ML) techniques has rendered DL-driven HAR increasingly prevalent. This study centers on the application of the EfficientNet transfer learning architecture with attention mechanism to the realm of HAR image recognition, specifically utilizing skeleton-based representations. The objective was to discern and evaluate the architecture’s performance attributes in this context. Notably, our results unveiled an impressive classification accuracy of 99% for EfficientNet. These findings underscore the efficacy of the EfficientNet model in enhancing HAR capabilities, showcasing its potential for pivotal roles in automated settings.