Real-Time Human Activity Recognition Using Convolutional Neural Network Methods and Deep Gated Recurrent Unit
Rangga Athallah Fajar, S-Y. Chou, Anindhita Dewabharata · 2023
This paper presents a real-time action recognition system for cleanroom standard operating procedures (SOPs). The objective is to develop a lightweight and efficient system capable of recognizing multiple actions and detecting individuals who deviate from the SOP. The proposed method utilizes a 3D Convolutional Neural Network for action classification. It employs object detection and tracking algorithms to focus on individuals performing the SOP. The proposed method can handle multi-object action recognition by incorporating object detection and tracking. The technique is designed to run in real-time on standard computers without hardware accelerators. Experimental results demonstrate that the proposed method achieves similar accuracy to MoViNets but with faster training and prediction times. Furthermore, the technique effectively handles multi-object action recognition and identifies individuals who skip parts of the SOP. The average prediction time of 0.03 seconds outperforms MoViNets' average prediction time of 0.05 seconds.