基于巡逻无人机的轻量型安全帽佩戴检测方法与应用

目前，安全帽检测系统主要使用固定摄像头，无法实现全区域检测，而基于深度学习的检测算法结构复杂、计算成本高，无法满足移动端和嵌入式设备的部署要求。针对上述问题，该文提出一种基于无人机的安全帽轻量型视觉检测算法。系统通过无人机平台搭载的相机对施工现场进行图像采集，并无线传输至后台计算机进行处理，检测算法基于 YOLOv5s 框架进行了轻量化改进。针对无人机采集影像中目标占比较小的问题，该文采用了多尺度检测、图像预处理、正负样本不均衡等方法，对 YOLOv5s 目标检测算法进行针对性改进。测试结果表明，与原模型相比，轻量型目标检测模型的平均精度均值仅下降了 1.72%，但在同一 CPU 上的推理速度提升了 1 倍，浮点计算量由原来的每秒 165 亿次压缩至每秒 34 亿次，模型大小约为原模型的 1/10。

Research on Safety Helmet Recognition Method and Application Using Patrol Unmanned Aerial Vehicle

The existing helmet detection system mainly uses a fixed camera, it cannot achieve full-area detection, and the previous detection algorithms based on deep learning have complex structures and high computational costs, which cannot meet the requirements of using mobile vehicles and embedded devices. In this paper, a lightweight helmet detection algorithm scheme based on unmanned aerial vehicle is proposed. The drone is loaded with camera to collect images of the construction site, and the image data is transferred to the computer via wireless communication. Based on the YOLOv5s target detection algorithm, a lightweight detection algorithm is investigated. To improve the detection more efficient, the YOLOv5s target detection algorithm is improved in terms of multi-scale detection, image preprocessing, unbalanced positive and negative samples, and inference speed. This design scheme combines deep learning and unmanned aerial vehicle technology, not only to realize real-time automatic detection of helmet wearing, but also can realize the full-area helmet detection of the construction site. Real experiments show that, the lightweight target detection model is only 1.72% lower than the mean average precision of the original model. The inference speed on the same CPU can be doubled, and the floating-point calculation is reduced from 16.5 billion to 3.4 billion times per second. The model size is almost 1/10 of the original size.