基于YOLOv10改进的车间人员不安全行为检测算法

Improved Unsafe Behavior Detection Algorithm for Workshop Personnel Based on YOLOv10

摘要: 针对车间、工地等生产作业场景中作业人员因未佩戴安全帽、吸烟、使用手机、未穿工服等不安全行为易引发严重安全事故的问题，本文提出一种基于YOLOv10改进的不安全行为检测模型。首先，引入小波池化模块重构YOLOv10的上下采样结构，以增强模型对细粒度特征的感知性能；其次，在主干网络中构建LSKA_SPPF模块，通过融合大核注意力机制提升多尺度特征提取能力；最后，设计遮挡感知检测头SEAMHead，通过显式建模通道与空间依赖关系，增强模型对遮挡区域的特征表达与补偿能力。在自建的WorkerBehavior数据集上的实验结果表明，改进后的模型在检测性能上优于基准YOLOv10，mAP@0.5提升了1.1个百分点，mAP@0.5:0.95提升了0.6个百分点，验证了所提方法在复杂工业场景下对车间人员不安全行为检测的有效性。

Abstract: Unsafe behaviors in production environments such as workshops and construction sites, including not wearing safety helmets, smoking, using mobile phones, and failing to wear work uniforms, can easily lead to serious safety accidents. To address this issue, this paper proposes an unsafe behavior detection model based on an improved YOLOv10. Firstly, a wavelet pooling module is introduced to reconstruct the up-sampling and down-sampling structure of YOLOv10, thereby enhancing the model’s ability to perceive fine-grained features. Secondly, an LSKA_SPPF module is constructed in the backbone network, which improves multi scale feature extraction capability by integrating a large kernel attention mechanism. Finally, an occlusion aware detection head named SEAMHead is designed. By explicitly modeling channel and spatial dependencies, the proposed head enhances the model’s feature representation and compensation capability for occluded regions. Experimental results on the self constructed WorkerBehavior dataset show that the improved model outperforms the baseline YOLOv10 in detection performance. The mAP@0.5 increases by 1.1 percentage points and the mAP@0.5:0.95 increases by 0.6 percentage points. These results demonstrate that the proposed method is effective for detecting unsafe worker behaviors in complex industrial scenarios.