Research On Driver Face Detection Algorithm In Monitoring Environment

At present,the mainstream highway monitoring system in China can complete the detection and feature extraction of driving vehicles,but because of the light and the complex background,the feature extraction of drivers’ faces inside the driving vehicles is not yet complete.Therefore,detecting the driver’s face under the monitoring environment will provide a critical help for the investigation of accident escape or other illegal acts.In this thesis,the MTCNN-based driver face detection algorithm is studied first,and the task of driver face detection is completed after targeted improvements.Then,the Yolo-based vehicle detection algorithm is studied,which limits the area of face detection to improve the accuracy of face detection through the detection of vehicles.Finally,with this method,the accuracy of face detection is improved.Two algorithms are the core of the design and development of a monitoring management software.The main contents of this thesis are as follows:(1)Driver face detection algorithm based on MTCNN.MTCNN is a three-level cascaded convolution neural network,which implements the task of face detection from coarse to fine.The following improvements are applied to solve the problems that exist in the driver driving image that the algorithm is applied in the monitoring environment: The feature point calibration task is introduced to improve the detection rate of driver’s face detection;A combination of suitable thresholds of prediction algorithm is selected through experiments to reduce the false detection rate of driver’s face detection;The idea of back propagation of difficult online samples is introduced for the model to improve the accuracy of model training and reduce the training time.(2)Yolo-based vehicle detection algorithm.Yolo is an end-to-end convolution neural network that can perform both regression and classification tasks.This algorithm is selected for vehicle detection to limit the area of driver face detection.The following improvements have been made to solve the problems that the algorithm can be applied to the driver driving image in the monitoring environment: The idea of a priori candidate box is introduced in the model,the shape and size of the border that appears most frequently in the training set are found by using the clustering algorithm,which improves the accuracy of the model to predict the vehicle frame.The idea of multi-scale training is introduced into the model to solve the problem of weak generalization ability when the same kind of object has an unusual aspect ratio.In order to ensure that the network can accept images of different sizes for training and prediction,the output layer of the network is fine-tuned,which improves the generalization ability of the model and ensures that the model can run robustly on images of different sizes.(3)Development of monitoring and management software to extract vehicle and driver face features.Based on the driver face detection algorithm and vehicle detection algorithm proposed in Chapters 2 and 3 as the core,a monitoring and management software is developed which can extract vehicle features and driver face features simultaneously.Starting from the design and implementation of the core module,the function of each module and the flow chart of the module operation are introduced respectively.