Hardhat Detection Using IR and Depth Frames

Abstract

The construction industry is one of the largest sectors and 20% of the total deaths happen here. Safety hardhat helmets are one of the precautionary methods recommended and followed in a construction site. Detecting and classifying people wearing and not wearing helmets is an important task. Safety and Security is one of the most researched fields in recent years and acquiring 3D geometric information from any real environment is an important task for such applications. Well-known methods like stereo vision camera systems suffer from high time consumption or from the inability to match corresponding points inhomogeneous regions. Time of Flight (ToF)technology, being a recent development, fulfills features desired for real-time distance acquisition along with the compact size and higher frame rate. A safety application based on the ToF technology ad IR imaging is proposed in the paper. The ToF sensor provides depth information for each pixel as opposed to RGB values in case stereo camera-based systems. The depth sense cameras provide IR images along with the depth information. A YOLO framework is used to classify images. YOLO being faster than RCNN and the faster RCNN is much suitable for real-time classification. The model was trained on 50,000 images. Weights obtained during the 6000th epoch were chosen. A Mean Average Precision of 56% was obtained while testing.