Research Of Driving Behavior Recognition Method Based On Visual Sequence Information Processing

Intelligent transportation is an important trend in current technological development,and autonomous driving has not yet achieved fully autonomous unmanned driving.Therefore,intelligent assisted driving system still has great practical value.Based on the real-time detection and recognition of the driver’s driving behavior,when the driver of the vehicle exsists driving behavior that does not meet the safety standards,the intelligent assisted driving system can issue an warning in time to remind the driver to correct the unsafe driving behavior.Therefore,taking advantage of intelligent assisted driving systems can effectively reduce the probability of traffic accidents.There are still many problems when the driving behavior method is actually applied to the intelligent assisted driving system,such as: repeated identification,high system energy consumption,privacy protection,and real-time model update.This thesis focuses on the problem of driving behavior recognition in the context of visual sequence information processing.In this thesis,driving behavior recognition is realized based on convolutional neural network,and driving behavior recognition method is practically applied based on federated learning and SSIM algorithm.The main research work of this thesis is as follows:（1）The visual sequence information is preprocessed based on the SSIM algorithm.Removing the driving behavior images with high similarity in the visual sequence based on similarity evaluation can solve the problem of repeated identification,and achieving the purpose of saving system energy consumption at the same time,which is useful for the driving behavior recognition method to be applied in embedded and other low-power devices.In terms of experiments,this thesis demonstrates the superiority of the visual sequence information preprocessing method based on the SSIM algorithm by designing comparative experiments.the similarity checking algorithm based on the SSIM algorithm is implemented and verified by experiments at the same time.（2）Training driving behavior recognition models based on federated learning.Federated learning can take advantages of the data resources of the existing vehicle platform to train models with excellent performance in real time while protecting the privacy of vehicle data,which is useful for the practical application of driving behavior recognition methods.In terms of experiments,the thesis designs comparative experiments from different training methods and different federated learning training parameters to test the impact of different training conditions on the training effect of the driving behavior recognition model.（3）Based on the SSIM algorithm and federated learning,driving behavior recognition method based on visual sequence information processing is realized.The application system is constructed from server to client and the preliminary experimental verification is accomplished.