Image analytics using Artificial Intelligence (Human Action Recognition in industrial workplace)

Abstract

The recent pandemic has reinforced the concept of industry 4.0 in traditional manufacturer industries, and one of the rising needs is to understand operators’ action to increase productivity and efficiency. Compared to traditional video action recognition tasks, video cation recognition under an industrial setting involves unusual objects, complex background and more inter-human interactions, which have an obvious gap between current public action recognition dataset. In this project, an industrial based dataset is being constructed to fill the blank in action recognition tasks in industrial workplace. Furthermore, two methods are proposed to improve the existing TSN and TSM model performance on human action recognition tasks via introducing the concept of grouping and split-attention mechanism to enhance model efficiency and accuracy. Various experiment setting and data augmentation methods are also reviewed in detail to explore the optimum setting in action recognition tasks. The model performance has improved from 78.80% to 90.51% on UCF101 dataset, and has reached 84.22% accuracy on self- constructed industrial dataset.