Research On Intelligent Recognition Of Bad Driving Behavior Images In Video On Urban Road Traffic

With the continuous improvement of the overall quality of life,people’s strong demand for cars has promoted the increase in car ownership year by year,making the frequency of road traffic accidents continue to increase.Especially for urban road traffic accidents,this has become an urgent problem in the process of urban development.Drivers’ bad driving behavior is the main cause of road accidents.Therefore,the intelligent recognition of the driver’s bad driving behavior images,especially the driving behavior images from the car,has become a main direction of current driving behavior research.Researchers have adopted different research methods to conduct a large number of studies,and obtained achievements of certain practical value.Among them,the heavyweight convolutional neural network can recognize driving behavior images well,but due to its large amount of parameters,it cannot be directly deployed in mobile devices with limited memory capacity and computing resources.Therefore,this article focuses on driving behavior video images,and conducts research on the recognition of lightweight convolutional neural networks on driving behavior images from the following two aspects:(1)Convolutional neural network based transfer learning for driving behavior image recognition.The four large parameter convolutional neural network models of VGG16,Res Net50,Inception V3,and Xception are fine-tuned using the deep transfer learning method on the AUC V2 driving behavior image dataset.The results on the test set indicate the performance of the Xception model.The best overall recognition accuracy rate is93.24%;the mainstream lightweight convolutional neural network model also uses the same method to conduct experiments on driving behavior image.The results showed that the Efficient Net-B0 model performs best,with an overall recognition accuracy rate of91.21% on the test set.In addition,by comparing the two models based on the confusion matrix and using the Grad-CAM feature visualization method,it can be seen that the Xception model is superior to the Efficient Net-B0 model in driving behavior image recognition.(2)The regularization method based on image data augmentation and knowledge distillation improve the recognition ability of Efficient Net-B0 model for driving behavior images.To address the problem that the Efficient Net-B0 model is inferior to the Xception model in the recognition of driving behavior images,three image data augmentation regularization methods and knowledge distillation are proposed to be used in the training process of the Efficient Net-B0 model.The experimental results show that the mixup data augmentation method and knowledge distillation can help improve the model’s recognition performance of driving behavior images;the proposed combination of the mixup data augmentation method and knowledge distillation is applied in the training process of the Efficient Net-B0 model.The experimental results show that when the Normal KD knowledge distillation method is combined with the mixup data augmentation method,the performance of the Efficient Net-B0 model on the driving behavior image can be effectively improved,and the overall recognition accuracy rate reaches 95.94% on the test set.In addition,by comparing and analyzing the improved performance of the Efficient Net-B0 model and the Xception model based on the confusion matrix and the Grad-CAM feature visualization method,it can be seen that the Efficient Net-B0 trained by the Normal KD knowledge distillation method combined with the mixup data augmentation method is better than the Xception model in image recognition of driving behavior.This article also developed an app based on the Android operating system to implement the performance-enhanced Efficient Net-B0 model for mobile deployment.