基于视频流的施工现场工人安全帽佩戴识别研究

王秋余

摘 要

佩戴安全帽是防止建筑工人头部损伤的有效方法之一。将计算机视觉方法应用于识别建筑工人的安全帽佩戴情况，能够加强对建筑工人的外部监督，从而减少头部损伤安全事故的发生率。然而，以往的方法通常依赖于样本的监督训练，存在单一图像遮挡性问题、小目标识别准确率低、不能适应场景的复杂环境等缺点。因此，本文针对施工现场视频特点与环境特点，提出了基于YOLO的半监督学习安全帽佩戴识别算法，并设计了安全帽佩戴识别系统。

首先，对YOLO网络结构改进得到建筑工人识别网络N_p与安全帽识别网络N_h；之后采用了公共数据集与真实环境图片数据集进行网络的预训练与离线训练，得到泛化模型。对其泛化模型采用半监督学习方式进行在线学习，进而提高算法在特定场景下的识别准确率与泛化能力；最后将基于YOLO的半监督学习安全帽佩戴识别算法移植到开发套件中，设计了安全帽佩戴识别系统。为了验证安全帽佩戴识别系统的识别性能，随机选取了某市地铁系统建设项目的视频流序列作为案例研究，数据得出在一般场景条件视频流的建筑工人与安全帽识别精度在85.7%至93.7%之间，显示该系统有较高的准确率。并在存在小目标对象与遮挡情况下，建筑工人与安全帽依旧有87.4%、86.2%的准确率。结果表明，安全帽佩戴识别系统能有效适用于复杂环境施工现场的安全帽佩戴识别工作，为建筑行业信息化的安全监测提供了新的研究视角与技术方法。

关键词：安全管理 在线学习 半监督学习 视频流 安全帽识别

Abstract

Wearing helmets is one of an effective way to prevent construction worker from head injuries. The computer vision method is applied to identify helmets worn by construction workers, which strengthens the external supervision of construction workers, and thus reducing the incidence of head-related safety accidents. However, there exists some problems in previous method based on the supervised training of the sample, like occlusion of single image, low accuracy of small target recognition, and inability to adapt to the complex environment of the scene, etc. Therefore, this paper proposes a semi-supervised learning safety helmet wearing recognition algorithm based on YOLO which matches video and environmental characteristics of construction site. Besides, a helmet wearing identification system is designed in this paper.

Firstly, a construction worker identification network and a helmet recognition network are established by improving the YOLO network. Secondly, these networks are pre-trained and offline trained using the public data set and the real environment image data set, and then the generalized model is obtained. After that, the generalized model starts to going on a semi-supervised learning. Which accordingly improves the algorithm's recognition accuracy and generalization ability in specific scenarios. Lastly, the helmet wearing identification system is designed by transplanting the semi-supervised learning helmet wearing recognition algorithm based on YOLO into the development kit. To verify the recognition performance of the helmet wearing identification system, the video stream sequence of the subway system construction project was randomly selected as a case study. The data shows that the identified accuracy of construction workers and helmets in the video stream sequence under general scene conditions is between 85.7. % and 93.7%, which shows that the system has a high accuracy and recall rate. The results show that the helmet wear identification system can be effectively applied to the helmet wearing and identification work in the complex environment construction site, and provides a new research perspective and technical method for the construction industry information security monitoring.

Key words: Safety Management Online-Learning Semi-Supervised Learning Video Streaming Hard Hat Recognition