A real-Time framework for dangerous behavior detection based on deep learning

Abstract

Identifying dangerous behaviors of workers in complex production operation scenarios is an important research field of intelligent monitoring technology. When implementing practical applications, it often has complex and diverse function requirements and business logic. To this end, this paper proposes a unified and simple framework for monitoring unsafe behavior based on deep learning technology. We first conduct data analysis and demand logic disassembly based on six actual production scenarios and decouple the complex task requirement as a coupled problem of object detection function, adaptive and variable scene recognition function, behavior analysis function, and safety logic reasoning function. Then we build a unified detection framework and use four sub-modules for real-Time and high-efficiency detection, which are a perceptron-based efficient scene recognition module, a Yolov5s-based real-Time object detection module, an area-based behavior judgment module, and a configurable safety rule inference module. While maintaining the characteristics of splitting and combining these modules for different application requirements, their effects on the test set have also reached the best. Among them, the accuracy rate of the scene recognition module has reached 100.00%, and the mAP of the object detection module has reached 99.06%, average FPS of the overall framework reached 72.05.