Enhanced Helmet Detection in Surveillance Systems with YOLOv6 for Accident Prevention and Safety Compliance

Abstract

Helmet detection is an essential aspect of achieving safety compliance and preventing accidents in dangerous situations like road traffic and construction sites. Safety compliance and prevention of accidents of high risk environments such as construction sites and traffic intersections is dependent on ensuring helmet usage. Yet, real time performance and accuracy of such systems, especially earlier versions of YOLO (e.g. YOLOv3, YOLOv5) face challenges in handling scenarios with occlusions, varying lightening conditions, and diverse helmets. The limitations of the above approaches are addressed in this paper by proposing a robust helmet detection framework based on the YOLOv6 architecture with proper transfers from synthetic to real-world surveillance. BiFPN performs advanced feature fusion to facilitate the detection of the helmet, the advanced composite loss uses CIoU and Focal Loss to improve localization and class balance, and introduces the concept of the novel post processing module—Helmet Geometry Validator (HGV)—that validates the detections using geometric shape feature to reduce the false positive from similar objects such as water bottle. Training and evaluation was performed on a diverse dataset of multiple environments. Additionally, the proposed model surpassed the baselines in terms of YOLOv3 and YOLOv4 as well as Faster R-CNN with precision of 89%, recall of 85%, F1-score of 87% and real time inference at 20 FPS. These results confer the viability of this proposed system and its promise of effectiveness for deployment in dynamic safety critical environments as well as provide a scalable, accurate and visual solution to automated helmet detection and compliance monitoring.