Industrial abnormal event detection using artificial intelligence

Abstract

Slips, trips and falls are abnormal events that prevalently contribute to workplace injuries in industrial areas. These injuries pose significant health risks and necessitate immediate intervention. Traditional fall detection systems often rely on wearable sensors, but they may be intrusive or inconvenient for continuous monitoring. Other forms of data monitoring methods presented can be computationally expensive. In this project, we propose a novel approach for detecting the abnormal event of slips, trips and falls using a lightweight skeletal pose estimator, combined with deep learning techniques. We extract skeletal pose data representing human body movements with the use of pose estimators. To analyse and classify these sequences, deep learning model such as the Long Short-Term Memory (LSTM) and Transformer networks were used. We evaluate the performance of the proposed fall detection system using cross-validation techniques and metrics such as precision, recall, and F1-score. Our results demonstrates that the use of skeletal pose features is indeed effective in the space of fall detection. This project contributes to the development of non-intrusive and robust fall detection systems by leveraging skeletal pose data and advanced deep learning techniques. The proposed approach holds promise for real-world deployment in industrial workplaces such as factories and construction worksites, ultimately enhancing the safety and well-being of individuals at risk of falls.