Research On Recognition And Early Warning Of Irregular Driving Behavior Based On Deep Learning

With the rapid development of the global economy,people’s living standards have been rising,and so has the ownership of motor vehicles.With this,traffic problems such as traffic congestion have arisen,the most serious of which are traffic accidents,causing huge loss of life and property.According to relevant information and related research,the driver’s irregular driving behavior is an important cause of traffic accidents,so timely identification and early warning of irregular driving behavior is of great significance.Irregular driving behavior includes fatigue driving,distracted driving,drunk driving,etc.The focus of this thesis is on the identification of fatigue driving and distracted driving.There is still room to improve the speed and accuracy of the existing algorithms for fatigue and distracted driving recognition.In this thesis,the two algorithms are optimized to improve the combined speed and accuracy performance of the model,and the main research is as follows:(1)The target detection model for fatigue driving recognition is improved to address the problems of slow speed and poor accuracy of fatigue driving recognition.The model can directly identify the driver’s eye-mouth state,simplify the fatigue driving recognition process,and improve the real-time recognition.The model is based on the YOLOv7-tiny network and uses the Ghost module to make lightweight improvements to further improve the realtime performance of the model,but with reduced accuracy.The original PANet structure is replaced with Bi FPN structure to enhance the multi-scale feature extraction capability of the network,which improves the accuracy of the network with almost no change in the number of network parameters and computational effort.The final model greatly improves the realtime performance of the recognition algorithm while ensuring high accuracy.Based on the results of this model,the fatigue recognition algorithm’s accuracy is improved by fusing eye and mouth features to jointly determine the driver’s fatigue status.(2)A target detection model for distracted driving recognition is improved to address the problems of slow speed and poor accuracy of distracted driving recognition.The model uses the Conv Ne Xt network as the backbone network and adopts the multi-scale feature fusion network architecture of PANet.In order to further enhance the feature extraction capability of the network and improve the network accuracy,while increasing the number of network parameters and computation as little as possible,a multi-headed self-attention mechanism is introduced at the end of the backbone network,where the feature map size is the smallest.The final model greatly improves the accuracy of the distracted driving recognition algorithm while ensuring high real-time performance.Finally,based on the two recognition algorithms proposed in this thesis,an irregular driving behavior recognition and early warning system was designed and implemented.The system was deployed to in-vehicle devices and was able to promptly identify and remind drivers to correct irregular driving behavior through voice,while uploading the records to the management platform.After the field test,it is proved that the recognition speed and accuracy of the system can meet the actual demand.