Research On Driver State Monitoring Algorithm Based On Multi-task Convolutional Neural Network

With the rapid development of the automobile industry in our country,traffic accidents have occurred more frequently,causing a large number of casualties and huge economic losses every year.Abnormal driving states such as driver fatigue are one of the major causes of traffic accidents.Driver status monitoring algorithms can reduce the probability of traffic accidents by issuing early warnings to the driver when the driver is in an abnormal driving state.In recent years,deep learning has been increasingly used in the field of image processing,including driver state monitoring related algorithms.However,few existing studies have optimized deep learning algorithms for driving conditions,resulting in a lack of rapidity and robustness.To address this problem,this paper proposes a driver status monitoring algorithm based on multi-task convolutional neural network.The network can benefit from temporal prior,improving detection accuracy significantly and bypass the step of face detection generally,leading to achieving a more rapid and accurate acquisition and calculation of the designed driver’s facial features,and classification of the driver state.The main contents of this research are as follows:1)Based on the SSD,an object detection network,this research optimizes the network structure and parameters for driving conditions and engineering processes,using the Relu6 activation function and L2 regularization loss to avoid variable expression problems such as truncation during deployment,so as to achieve high efficiency and accuracy in face detection on edge computing devices.The runtime of the optimized face detection network in the experiment is in line with expectations,and the accuracy is also satisfying.2)We propose a multi-task convolutional neural network to accomplish facial landmarks detection,head pose estimation and face verification simultaneously,which can make use of temporal prior to improve accuracy of its tasks.This paper specially designs the form of the input data and training technics.A brand-new prior knowledge generator is designed to transfer prior knowledge in the form of he at maps.Experiments show that temporal prior knowledge can greatly improve the accuracy of the network’s tasks.The network has the characteristics of extremely low runtime and high accuracy,and can meet the requirements of the driver state monitoring algorithm.3)This paper designs a number of driver facial features and their corresponding extraction and calculation methods.An SVM classifier is designed based on these features.We also collect and produce a small-scale abnormal driving dataset to verify the efficacy of the algorithm.Experiments demonstrate that the trained SVM classifier can complete the driver state classification task accurately in a short time in the driver state monitoring algorithm in this paper,which proves the efficacy of the driver state monitoring algorithm proposed by this paper.The multi-task convolutional neural network using temp oral priors proposed in this paper has 6.75% and 2.25° errors in facial detection and head pose estimation without prior knowledge.Valid prior knowledge can benefit the network,reducing errors of facial detection by 27.21%.The driver state classification accuracy is 90.28%.The algorithm in this paper has remarkable real-time performance and high accuracy,which has significant practical value.