基于YOLOv4-Tiny与显著性检测的安全帽佩戴检测

Helmet Wearing Detection Based on YOLOv4-Tiny and Saliency Detection

传统的采用人工巡视工作场所或使用实时监控,对工人安全帽佩戴情况进行检测的方法费时费力,效率低下。随着深度学习技术的发展,可以利用目标检测的方法来对工人安全帽佩戴进行实时检测。但受工作场所复杂环境因素的影响,直接应用目标检测模型检测工人佩戴安全帽的准确率不高。为此,通过优化YOLOv4-Tiny模型,利用K-means聚类算法,针对安全帽这个特定目标生成合适的先验框,然后将YOLOv4-Tiny检测结果与改进的动态显著性检测结果有效融合,提升了工人佩戴安全帽检测的准确率。实验证明,该方法能有效缓解复杂背景对目标检测的影响,能够较准确地检测出有无佩戴安全帽的工作人员。

Reviewing of the wearing of helmets by workers in traditional way is time-consuming,laborious and inefficient.With the development of deep learning,the object detection can be used to detect the situation of workers wearing helmets in the factory floor.But object detection model will be affected by the factory's complex environment.It will also decrease the accuracy of the model if the object detection model is used directly.In order to solve these problems,this paper improves YOLOv4-Tiny by using K-means clustering to make the prior boxes suitable for the size of helmet.Then the results of object detection compared with the results of dynamic saliency to recheck the detection target.It will improve overall detection accuracy.The experiment proves that this detection method will decrease the influence of complex background in object detection.And it can identify the staff who wearing the helmet or not accurately.