基于YOLOv4算法的骑乘人员头盔佩戴的远程检测

Remote detection of helmets for riders based on YOLOv4

针对现有头盔检测模型易将非骑行者纳入检测范围的错误识别问题,提出将人与车整体标定的方法来制作骑行者佩戴头盔数据集,然后采用YOLOv4算法进行模型训练,可以正确识别出骑行状态与非骑行状态。通过二分K-means算法对拟识别目标先验框进行聚类分析,采用分步训练方式优化学习权重,识别精度达到了90.8%。其次,为了将交通执法人员从危险环境解放出来,提出网络摄像头截取视频流并通过以太网输入到PC端YOLOv4模型中的方法,替代人工实现远程检测。实验表明,该方案能正确识别出视频中骑行人员有无佩戴头盔,证明了该方案的有效性。

Traffic in cities is complex,which causes traffic accidents easy to happen.When motorcycle riders wear helmets during riding,their accidental injuries can be reduced to a certain extent.Therefore,it is of great significance to carry out remote detection on whether riders wear helmets or not during riding.Aiming at the erroneous identification problem that exists in current helmet detection models which tend to include non-cyclists in the detection range,the model in this paper proposed the method of overall calibration of taking people and motorcycles as a whole to produce helmet wearing data sets of cyclists,and then YOLOv4 algorithm was used for model training,so that the cycling state and non-cycling state could be correctly identified.The binary K-means algorithm was used to perform clustering analysis on the priori frame of the target to be recognized,and the learning weight was optimized by step training method,and the recognition accuracy reached 90.8%.Secondly,in order to liberate the traffic law enforcement personnel from the dangerous environment,a method was proposed to capture the video stream by webcam and input it to YOLOv4 model at PC terminal through Ethernet instead of manual so that remote detection could be achieved.Experimental results showed that the scheme could correctly identify whether the cyclists were wearing helmets in the video,which proved the effectiveness of the scheme.