A pull-reporting approach for floor opening detection using deep-learning on embedded devices

The construction site is prone to a considerable number of accidents due to its dense and complex nature. Accidents due to falling from opening at the construction site are leading reasons for severe injuries and sometimes fatalities. Openings and Holes are made on floors and roofs during the building construction or destruction. Despite being aware of the hazards associated with openings, gaps, and holes when working at heights, many workers fail to cover the openings or remove the covers for ease of work. The current inspection procedure relies on manual practices that are error-prone, time-consuming, expensive, and difficult for a site manager to monitor. Therefore, the authors propose a pull-reporting approach to resolve this issue by utilizing a computer vision detector model embedded in the android mobile to facilitate the safety manager. The proposed application uses YOLOv4 trained weights on a custom dataset obtained from the recorded videos at the Korean Scaffolding Institute with various view angels and various degrees of occlusion and data crawling techniques. The weights are then deployed on edge devices using TensorFlow API, Java programming, and maintain a real-time database of unsafe behavior. The developed system can identify, classify, and record the fully opened openings (FOO) and partially opened openings (POO) at the construction site along with geo-coordinates details in real-time.