Research And Implementation Of Driver Abnormal Behavior Detection System Based On Deep Learning

In recent years,with the rapid development of China’s automotive electronics industry,the total number of motor vehicles in China is increasing significantly.However,with the popularity of motor vehicles,people’s standard of living is facilitated while traffic accidents can also occur.Therefore it is of great importance to identify drivers’ behaviour in real time and to alert them to dangerous driving behaviour.Therefore,this paper combines the practical aspects of improving the accuracy and real time of driver behaviour recognition algorithms,focusing on face detection and face feature point localisation,to improve and optimise the driver abnormal behaviour detection algorithm,and proposes a driver abnormal behaviour detection method.Firstly,we did model training for the multi-task cascaded MTCNN network and analysed its experimental results,and then improved and optimised the network structure and parameters of the three sub-networks P-Net,R-Net and O-Net of MTCNN respectively.After the experiments,it was shown that the accuracy of the MTCNN network improved from 96.2% to 97.7% after the optimization,and the accuracy of the network improved by 1.5%.After the detection of the face frame is obtained,this paper chooses to use the ERT cascade algorithm to finally achieve the precise localisation of 68 key points of the face.Then,a face tracking method based on the fusion of TLD and SAMF is proposed.The tracking part of the TLD algorithm is replaced with a tracking detection and SAMF algorithm suitable for TLD,and the HOG features are fused into the detection part of TLD.Simulation experiments are done using the face tracking algorithm with MTCNN network + TLD + SAMF fusion,and according to the experimental results,the proposed algorithm has a very good detection effect.Then,we analyze and extract the driver’s eyes,mouth and head features.For extracting eye features,the EAR eye aspect ratio algorithm is used,and the method of using the average of both eyes is proposed to calculate the EAR value;for mouth feature extraction,the MAR algorithm is used,and the method based on the inner contour of the mouth is proposed to calculate the value of MAR;for extracting head posture features,the pitch angle Pitch in 3D space is used to determine their head posture,and the estimation of distracted driving behavior is established model;for the recognition of the behavior of answering phone calls and the localization of the ear region,the skin color segmentation method of YCb Cr and the continuous time monitoring are used to complete the detection in the daytime situation,and the maximum contour finding method is proposed to complete the detection in the nighttime situation.Finally,the driver abnormal behavior detection system was designed using Open CV and MFC technologies.The detection performance of the system is reasonably evaluated by experimentally verifying each detection module separately,and it is proved that the designed driver abnormal behavior detection system not only has high detection accuracy but also has good real-time performance.Its detection accuracy is above 90%,and its real-time performance can reach more than 10 times per second.