基于YOLOv3改进算法的出租车驾驶人口罩检测

Detection of taxi driver's mask wearing based on improved YOLOv3 algorithm

为了更加及时精准地监测出租车驾驶人在工作状态下的口罩佩戴情况,提出一种基于YOLOv3改进算法的出租车驾驶人口罩佩戴检测方法。在试验开始前,利用收集到的出租车车内视频数据制作一套包含2 478张出租车内部驾驶人工作状态下的图片数据集,根据数据集的特点采取3种改进策略:首先将主体网络中的普通卷积替换为深度可分离卷积,降低模型参数量的同时实现模型压缩和网络结构深度的增加;然后为了保证口罩边缘信息在多尺度预测过程中的融合效果,将原始算法中在3个特征图进行多尺度融合的策略减少为2个;最后为了维持在特征图融合过程中的锚框数量,采用了K均值(K-means)算法重新计算出8个初始锚框值,给2个融合特征图上分别分配4个初始锚,通过以上改进使算法能够更好地适配于自建数据集。研究结果表明:通过改进后的YOLOv3算法在驾驶人口罩佩戴检测时精度可以提升至96.2%,且模型被压缩到32 M,在英伟达1080Ti环境下处理速度为43帧/s,满足实时性需求要,改进后的算法表现更加优异,可以有效地用于出租车驾驶人口罩佩戴检测。

In order to monitor the mask wearing of taxi drivers in working state more timely and accurately,a mask wearing detection method for taxi drivers based on YOLOv3 improved algorithm was proposed.Before the beginning of the experiment,using the video data collected from the previous work,a dataset containing 2478 images of taxi drivers working inside was created.According to the characteristics of the dataset,three improvement strategies were adopted.First of all,the ordinary convolution in the main network was replaced by the deep separable convolution,the model compression and the increase of network structure depth were realized while reducing the number of model parameters.Then,in order to ensure the fusion effect of mask edge information in the multi-scale prediction process,the original three feature images was reduced to two in the multi-scale fusion.Finally,in order to maintain the number of anchor frames in the feature image fusion process,K-means algorithm was used to recalculate 8 initial anchor frame values,and four were allocated on each fusion feature image,through the above improvements,the algorithm can better adapt to self-built datasets.The results show that the accuracy of the improved YOLOV3 algorithm can be improved to 96.2%,and the model is compressed to 32 M.The processing speed is 43 frames per second in NVIDIA 1080Ti environment,which meets the real-time requirements.Therefore,it can be seen that the improved algorithm performs better and can be effectively used for mask wearing detection of taxi drivers.2 tabs,10 figs,23 refs.