Research On Driver Face Detection Technology Facing Traffic Monitoring

The traffic monitoring system in China can automatically capture the vehicles passing the traffic bayonet and extract relevant features,but the recognition of the driver's face is not covered yet.Although vehicles can be uniquely identified by license plates,some criminals usually use decks to commit crimes.This undoubtedly increases the difficulty of investigation.Under the background of the identity authentication of relevance to improve,the driver faces will gradually become an important vehicle-related feature and assist in the investigation on illegal vehicles.Therefore the driver's face feature extraction is of great practical significance.This paper focuses on images collected by the bayonet,and firstly,the most representative Adaboost algorithm in traditional machine learning algorithm is studied and improved.On the basis of this,the MTCNN face detection algorithm based on deep learning framework is further studied and effectively solve the bayonet images in the light changes,occlusion and other issues.The main contents of this paper are as follows:(1)Adaboost face detection algorithm.In view of the bayonet image environment,the training samples are collected,and the face classifier is trained.In training the classifier,the weight of the difficult samples will increase continuously,the classifier will be degraded.Therefore,the weight update strategy of the sample is improved.In order to reduce the false detection rate,a negative sample expansion strategy is proposed,which reduces the false detection rate to a certain extent.Finally,the Gaussian skin color model is used to screen candidate face areas.(2)MTCNN face detection algorithm.MTCNN is a three-stage cascaded convolutional network.A learning strategy of online difficult samples is proposed in order to decrease the model's training time,which can effectively speed up the error convergence during training.The idea of this face detection framework originates from CascadeCNN.Compared with CascadeCNN,MTCNN framework adjusts the size of convolution kernel in convolution layer from 5 * 5 to 3 * 3,and coalesces face detection task and border regression task into the same network.Experiments show that the network structure improve the detection rate and save the test time.Finally,the convolutional network has a good detection effect on images of color distortion and the dramatic changes in lighting conditions,showing a strong robustness to changes in the environment.(3)Design and development of driver face feature extraction and vehicle characteristics management.Firstly,we design the system from system performance and software architecture.Secondly,we show the implement details of core modules,including face detection and facial point detection.At last,we use C++ to achieve core code,and design MFC software interface for interactive display.