非机动电动车骑行人头盔佩戴检测研究

Helmet Wearing Detection of Non-motorized Electric Vehicle Riders

随着我国城市交通的不断发展,非机动电动车的数量不断上升,绿色出行的同时,交通事故也日益增多。为了保证非机动电动车骑行人在道路上的安全,国家逐步制定法规,要求出行需佩戴头盔,各地交通部门执行检查与处罚,对骑行人佩戴头盔的自动检测也提上了日程。提出用改进YOLOv5算法视频识别头盔佩戴情况,通过自组织非机动电动车骑行人头盔佩戴情况的数据集,选取聚类算法修改初始锚定框参数。然后,改进算法来适应样本集合,利用聚合输出网络提高判别的准确率。其中,采用迁移学习来减少训练资源的消耗。经各种场景实验测试的结果表明,在每秒30帧视频流下,检测的均值平均精度(mean average precision,mAP)达到了94.53%,满足对头盔佩戴检测精度和速度的要求。

With the continuous development of urban transportation in my country,the number of non-motorized electric vehicles is rising,and the number of traffic accidents caused by green travel is also increasing.In order to ensure the safety of non-motorized electric vehicle cyclists on the road,the state has gradually formulated regulations requiring helmets to be worn when traveling,and local traffic departments have implemented inspections and penalties,then automatic detection of cyclists wearing helmets has been put on the agenda.It is proposed to use the improved YOLOv5 algorithm for video recognition of helmet wearing.Through the data set of self-organized non-motorized electric vehicle cyclists’helmet wearing situation,a clustering algorithm is selected to modify the initial anchor frame parameters.Then,the algorithm is improved to adapt to the sample set,and the discriminant accuracy is improved by using the aggregated output network.Among them,transfer learning is adopted to reduce the consumption of training resources.The results of experimental tests in various scenarios show that under the video stream of 30 fps,the mean average precision(mAP)of detection reaches 94.53%,which meets the accuracy of helmet wearing detection and speed requirements.