Multi-Scale Safety Hardhat Wearing Detection using Deep Learning: A Top-Down and Bottom-Up Module

Abstract

Construction sites are the most unsafe and risky places where thousands of workers are injured and die every year throughout the world. Some protective gear like hardhat can protect personnel from unexpected accidents. Administrators need to confirm all personnel put on hardhat on their heads during working time. However, it is inefficient and time-consuming to monitor this task manually. Hence, an automatic system may give convenience to detect personnel whether they wearing hardhat or not when they are on duty. RatinaNet is used to detect and localize the hardhat/head of personnel into the construction site. ResNet50+Feature Pyramid Network (FPN) is used as the backbone of the architecture, a classification and a regression sun-module are used to classifying objects and localizing bounding box around the object. A robust semantical description is achieved using both top-down pathways and lateral connections. Hardhats or heads are detected on a multiscale using the bottom-up and top-down modules. Experimental analysis on a dataset using RatinaNet produces a prominent result that may be usable in real-time applications.