International Research Journal of Engineering and Technology (IRJET)
e-ISSN: 2395-0056
Volume: 13 Issue: 06 | Jun 2026
p-ISSN: 2395-0072
www.irjet.net
Safety Helmet Detection Based on Improved YOLOv8n-SLIM-CA Ravina Suresh Edake1, Dr. Niranjan Tukaram Kulkarni2 1Student, Alamuri Ratnamala Institute of Engineering and Technology (ARMIET), Maharashtra, India 2Associate Professor, New Horizon Institute of Technology and Management (NHITM)
---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - Ensuring worker safety in construction and
tired, responses to violations are slow, and you’re stuck with subjective error-prone assessments.
industrial settings always comes down to proper monitoring, and making sure everyone actually wears their safety helmets is a huge part of that. But traditional manual supervision is not only exhausting for the people stuck watching footage all day; it’s also error-prone and just not practical when it comes to large-scale, round-the-clock monitoring. Deep learning and object detection have changed the game for real-time safety monitoring. YOLO’s speed and accuracy set a standard, but tracking small, far-off, or partly hidden helmets on chaotic work sites? That still remains a significant challenge for existing systems. This research introduces an upgraded safety helmet detection system built on a customized YOLOv8 base called YOLOv8n-SLIM-CA. Here’s what makes it tick: we use mosaic data augmentation to boost the network’s sensitivity to small objects and make it generalize better, even if the work site always looks different. Coordinate Attention (CA) gets baked into the backbone network, sharpening the model’s focus on helmet regions and dampening visual distractions in the background. We slim down the neck architecture for lean, speedy multi-scale feature fusion ideal for lower-power, realtime applications. On top of that, we drop in a dedicated small target detection layer to crank up accuracy for far-off workers; you spot more helmets, even at a distance or in a crowd. We test the model with all the standard metrics: precision, recall, and mean Average Precision (mAP). Results show pretty clearly detection improves for helmets, even in complicated, high-clutter environments, all while keeping things fast enough for live deployment. Since this build has a tight computational footprint, it plays well with edge devices and embedded hardware. With YOLOv8n-SLIM-CA, we offer a smarter, lighter way to automate safety checks, so compliance goes up and risk goes down.
As artificial intelligence pushes deeper into safety solutions, computer vision models especially deep learningbased object detectors are making it possible to keep eyes on every worker at all times, without human fatigue. The YOLO architecture has built a reputation for balancing high accuracy and speed, and the latest iteration, YOLOv8, adds even better feature extraction and fast inference. Even then, recognizing helmets out in the wild is tricky. Open construction zones look nothing like clean test images there are messy backgrounds, weird lighting, and too many people in frame. Helmets end up small or blurred by distance, and overlapping workers or equipment hide them further. All this means YOLO models (and their competitors) often call out helmets where there are none, or worse, completely miss real ones. To solve these issues, simply stacking more layers or making the network deeper doesn’t cut it models grow too bulky and slow for edge devices. Instead, injecting smarter modules like channel and spatial attention, plus streamlining multi-scale feature fusion, can make a model sharper and more robust without bogging down processing. That’s where YOLOv8n-SLIM-CA comes in. By adding mosaic augmentation, coordinate attention, a slimmed-down neck, and a small-object detection layer, this version nails down the details that legacy systems miss and runs light enough for real-time, on-site deployment.
2. LITERATURE REVIEW
Key Words: Safety Helmet Detection, YOLOv8, Deep Learning, Object Detection, Coordinate Attention, Small Target Detection, Computer Vision, Real-Time Monitoring, Industrial Safety, Edge Computing
N. Fatima and her team (2025) [1] bolstered YOLOv8n with Coordinate Attention and a Slim-Neck structure. By highlighting meaningful image regions and discarding noisy features, they made the detection of tiny, far-away helmets more reliable. Slim-Neck stitched together multi-scale features without fluff, but, admittedly, the upgrades added to model complexity possibly making lightweight deployment more difficult.
1. INTRODUCTION Worker safety for industries like construction and mining usually centers around rigid rules for wearing protective gear especially helmets. Head injuries are no joke; they’re a leading cause of serious accidents mostly because people cut corners or the boss misses someone not following protocol. The standard way? Watching security videos manually. That just sets up a bunch of predictable problems: supervisors get
© 2026, IRJET
|
Impact Factor value: 8.315
Then there’s the FGP-YOLOv8n model from L. Zhang (2025) [2], which swapped out the usual backbone for FasterNet and slotted in lightweight attention layers. The result? The model not only trimmed down resource demands but still lifted mAP. The catch is, by altering the backbone so
|
ISO 9001:2008 Certified Journal
|
Page 864