基于SSD-MobileNet的安全帽检测算法研究

Research on the Algorithm of Helmet Detection Based on SSD MobileNet

文章提出以单发多盒探测器为基础检测框架,MobileNet为特征提取网络,在人体区域内检测安全帽,根据两个检测模型的输出结果判断安全帽是否被正确佩戴。根据测试结果,人体检测模型针对人体的检测精确率达到91.52%,召回率达到89.25%;安全帽分类检测模型针对安全帽的检测精确率达到88.32%,召回率达到85.08%;针对头部的检测精确率达到88.02%,召回率达到86.02%。在真实环境中对本文提出的检测方法的检测效果进行验证,相比传统的单发多盒探测器检测方法,平均精确率均值上升了2.79%,模型体积缩小为传统单发多盒探测器检测方法的五分之一,检测速度也提升了两倍。

This paper proposes a detection framework based on single shot multibox detector(SSD) and a feature extraction network based on mobilenet to detect the helmet in the human body area, and judge whether the helmet is worn correctly according to the output results of the two detection models. According to the test results, the detection accuracy rate of the human detection model for the human body is 91.52%, and the recall rate is 89.25%;the detection accuracy rate of the helmet classification detection model for the helmet is 88.32%, and the recall rate is 85.08%;the detection accuracy rate for the head is 88.02%, and the recall rate is 86.02%. Compared with the traditional SSD detection method, the average precision(map) of the proposed detection method is increased by 2.79%, the model size is reduced to one fifth of the traditional SSD detection method, and the detection speed is also increased by two times.