Helmet detection in traffic scenarios: enhanced performance for complex environments

In the cutting-edge development of computer vision technology, the high missed detection rate caused by target occlusion and the high computational cost of model inference remain core technical bottlenecks restricting the deployment of robust object detection systems in real-world scenarios. This paper presents an improved YOLOv10n-based algorithm to address these issues. Adaptive-DySample (ADS) enhances target detection via three innovative mechanisms-dynamic alignment, challenging sample prioritization, and multi-scale fusion with range adjustment-enabling flexible, object-aware upsampling to overcome traditional limitations. C2fCIB-Fusion(C2fCIB-F), built on the original C2fCIB, systematically addresses the shortcomings of traditional modules in terms of feature extraction flexibility, multi-scale adaptability, and inference efficiency through a series of methods: integrating multi-scale and dynamic channel optimization, introducing Spatial-Channel Interaction, adopting RepVGGDW, and utilizing residual connections. Additionally, Gated Multi-Scale Fusion Convolution (GMFConv) uses a dual-branch architecture (one for global features, one for local details) with a gating mechanism for feature fusion, enhancing small object detection capability and being lightweight for resource-constrained scenarios. Experimental results show the improved model achieves 2.4% higher mAP@0.5 and 5.1% better parameter than the original. It significantly enhances helmet detection accuracy under complex scenarios, maintains real-time performance, and balances a lightweight design, suitable for edge computing and safety monitoring systems.