Unsafe-Net: YOLO v4 and ConvLSTM based computer vision system for real-time detection of unsafe behaviours in workplace

作     者：Önal, Oğuzhan Dandıl, Emre 

作者机构：Department of Electronic and Automation Vocational School Bilecik Seyh Edebali University Bilecik Turkey Department of Computer Engineering Faculty of Engineering Bilecik Seyh Edebali University Bilecik Turkey 

出 版 物：《Multimedia Tools and Applications》 (Multimedia Tools Appl)

年 卷 期：2024年

页      面：1-27页

核心收录：

学科分类：08[工学] 0710[理学-生物学] 0810[工学-信息与通信工程] 1004[医学-公共卫生与预防医学(可授医学、理学学位)] 1002[医学-临床医学] 070207[理学-光学] 1001[医学-基础医学(可授医学、理学学位)] 081203[工学-计算机应用技术] 0837[工学-安全科学与工程] 0835[工学-软件工程] 0803[工学-光学工程] 0836[工学-生物工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 0702[理学-物理学] 

基　　金：Open access funding provided by the Scientific and Technological Research Council of T\u00FCrkiye (T\u00DCB\u0130TAK). This study was financially supported by the Scientific Research Projects Coordinatorship of Bilecik Seyh Edebali University with project number 2019-02.B\u015EE\u00DC.01\u201303 

主　　题：Object detection 

摘      要：Unsafe behaviour is a leading cause of death or injury in the workplace, including many accidents. Despite regular safety inspections in workplaces, many accidents occur as a result of breaches of occupational health and safety protocols. In these environments, despite efforts to prevent accidents and losses in hazardous environments, human error cannot be completely eliminated. In particular, in computer-based solutions, automated behaviour detection has low accuracy, is very costly, not real-time and requires a lot of time. In this study, we propose Unsafe-Net, a hybrid computer vision approach using deep learning models for real-time classification of unsafe behaviours in workplace. For the Unsafe-Net, a dataset is first specifically created by capturing 39 days of video footage from a factory. Using this dataset, YOLO v4 and ConvLSTM methods are combined for object detection and video understanding to achieve fast and accurate results. In the experimental studies, the classification accuracy of unsafe behaviours using the proposed Unsafe-Net method is 95.81% and the average time for action recognition from videos is 0.14 s. In addition, the Unsafe-Net has increased the real-time detection speed by reducing the average video duration to 1.87 s. In addition, the system is installed in a real-time working environment in the factory and employees are immediately alerted by the system, both audibly and visually, when unsafe behaviour occurs. As a result of the installation of the system in the factory environment, it has been determined that the recurrence rate of unsafe behaviour has been reduced by approximately 75%. © The Author(s) 2024.