FD-YOLOv5: A Fuzzy Image Enhancement Based Robust Object Detection Model for Safety Helmet Detection

Numerous deaths are being reported at sites where safety helmets are important for workers. Though these sites are usually equipped with closed circuit television (CCTV) cameras, real-time monitoring of these feeds proves to be a mammoth task as it requires large human personnel leading to financial burdens. This work harnesses the object detection abilities of various deep learning models to propose a robust solution for safety helmet detection problem. This work divides the overall task into three sub problems, where the first two sub problems explore various object detection models including Single Shot Detector (SSD), Faster Region Based Convolutional Neural Networks (FRCNN), YOLOv3, YOLOv4, and YOLOv5 over multiple benchmark datasets. In Sub Problem 1, the YOLOv5 model achieved a mAP of 93.5%, 94.1% and 92.7% on the three selected datasets, while it achieved a mAP of 93.8% over the selected dataset of the Sub Problem 2. In both these sub problems, the YOLOv5 model outperformed all the recently proposed models for helmet detection task and hence was selected as the model for further experiments. In the final sub problem, a novel fusion of YOLOv5 and a fuzzy based image enhancement module, was further fine-tuned to create a robust model (FD-YOLOv5 M) that worked visibly better than simple YOLOv5 even over real site noisy CCTV feeds. The proposed FD-YOLOv5 model, when tried over an enhanced dataset, efficiently detected safety helmets worn by humans and even differentiated them from various types of other common headgears. These observations strongly support the application of the proposed model at real site scenarios to detect workers wearing safety helmets.