Research On Driver Face Detection Method In Traffic Monitoring Based On Convolutional Neural Network

In recent years,with the number of cars increasing,the pressure on traffic is becoming greater,and the problems of traffic safety have increased,especially illegal behaviors such as hit-and-run and vehicle deck have occurred from time to time.In the system of traffic monitoring,the traffic police can extract vehicle information from the monitoring equipment of each road crossing and analyze the information to effectively combat these illegal actions.Among them,the information of driver face is the direct and effective feature information,which can help the traffic police to complete the tasks of traffic management.Therefore,it is of great practical significance to detect the driver face to obtain the face feature information.However,affected by the complex driving environment,the pictures of driver face often have the problems of insignificant local features of the face and the complex illumination variations,which make the detection of driver face face challenges.In this paper,we focus on these two issues to research the driver face detection.The main contents are mainly divided into the following aspects:(1)Focusing on the problem that the local features of the driver face are not significant,this paper designs a cascaded convolutional neural network(Driver Net).The Driver Net consists of three light convolutional neural networks with smaller convolution kernel and introduces the maximization feature map activation function(MFM).This method makes the network pay more attention to the study of driver face details and get more discriminative features of driver face.In addition,the cascade structure of layer-by-layer screening also minimizes the false detections of driver face.(2)Focusing on the problem of complex illumination variations in driver face picture,firstly,the data enhancement method based on Gamma correction is used to enhance the diversity of illumination variations in driver face.In particular,the driver face image with normal illumination is processed by complex illumination on the areas of driver face under the Io U.Secondly,by using feature fusion method in the third-level convolutional neural network of Driver Net,the middle-level features of the convolutional neural network containing more geometric information of image scene space are fused with the final high-level features,which improves the detection performance of the network for driver face pictures under complex illumination variations.(3)Focusing on the hard samples of the driver face pictures that have the problems of insignificant local features and complex lighting variations,the paper proposes an online hard sample mining method based on focus loss(FL-OHEM).This method makes the network pay more attention to learning the hard samples of driver face by real-time mining the hard samples and increasing the learning rate of the hard samples which are difficult to detect in the driver face samples during the training process of the convolutional neural network.This method effectively improves the performance of the detection of driver face by the network.