Research On The Method Of Driver Abnormal Posture Recognition Based On Convolutional Neural Network

The rapid development of urban transportation has brought great convenience to human life,but the occurrence of traffic accidents has brought endless troubles and disasters to human beings,which has seriously damaged the stability and prosperity of society.According to statistics,traffic accidents caused by abnormal driving of motor vehicles account for more than 90% of all traffic accidents.When the driver is driving,he or she is in an abnormal driving posture such as making phone calls,drinking water,or smoking,which will seriously affect the effective control of the driver and cause a traffic accident.Therefore,it is extremely necessary to monitor the driver’s abnormal driving posture and intervene to reduce the incidence of traffic accidents.The current driver monitoring system solutions are single classifier solutions based on shallow machine learning algorithms and deep learning algorithms.However,such solutions cannot effectively resolve the contradiction between real-time and accuracy.Based on this background,this paper studies and implements driver pose recognition method based on human pose estimation and convolutional neural network.Firstly,based on the Openpose posture estimation algorithm,this paper establishes a driver posture detection model and uses the model to detect the driver’s upper body joint points.The average detection accuracy in the homemade driver posture test set is 94.76%.In order to improve the detection efficiency of the posture classifier and improve the real-time performance of the system,this paper improves the original Openpose network structure based on some ideas of the Lightweight-Openpose posture estimation algorithm,and establishes a new Openpose-2 posture detection model.The test set has a detection accuracy of 91.44%,which can be run under a pure CPU,with an average FPS of 28.Secondly,based on the driver posture detection model,a criterion for judging the driver’s abnormal posture is proposed.The human joint point detection model is used to extract the coordinate information of each joint point of the driver in normal driving posture and abnormal driving posture,and the Gaussian mixture model EM algorithm is used to cluster the coordinate data to study the driver’s limb position in normal and abnormal posture.The rule is to determine whether the driver is in an abnormal posture.The experimental results show that the proposed criteria for identifying abnormal driving posture is simple and efficient,and the average recognition accuracy on the test set can reach 97.46%.Thirdly,in order to improve the practicability of the driver monitoring system,after determining whether the driver is in an abnormal driving posture,it is necessary to further identify the specific type of abnormal driver posture.This paper builds driving behavior classification models based on four network architectures: VGG-19,Inception-V3,Resnet-50,and Densenet-121,and compares and analyzes their comprehensive performance in driving posture recognition tasks.The results show that the behavioral classifier trained based on the Inception-V3 network architecture has the best overall performance,the recognition accuracy on the test set can reach 93.32%,and the FPS can reach 12 when detecting video under the CPU.Finally,a new driver abnormal posture detection system solution is proposed and implemented to solve the problems of current single classification model with low detection efficiency and serious computing resource occupation.This system combines the actual driving conditions and system performance requirements of the driver,and integrates two models of posture detector and behavior classifier,which can determine whether the driver is in normal or abnormal driving posture according to the abnormal driving posture discrimination criteria.Subdivision based on behavior classifier.The results show that the system achieves an average recognition accuracy of more than 90% when performing driver gesture recognition tasks,which can increase the detection efficiency by up to 2.33 times compared to a single classification model scheme.