Research And Application Of Federated Learning Method For Driver Distraction Recognition In V2X

Since a large number of traffic accidents are related to the driver distraction,the driver distraction recognition system can remind the driver in time when the driver makes a distracted action,which could effectively avoid the occurrence of traffic accidents.In recent years,the research on driver distraction recognition has mainly focused on improving the accuracy of the recognition model,however,this thesis focuses on the training and deployment manner of it.Different from traditional training and deployment ones,this thesis adopts edge training and local deployment manner,that is,each vehicle node uses the edge computing method of federated learning to complete model training,and the trained model is then deployed locally for driver distraction recognition,which has the advantages of good real-time performance,easy model upgrade and privacy protection.At present,federated learning has yet to be explored in terms of reducing the impact of non-iid data and communication overhead,and there are few researches on applying federated learning to the driver distraction recognition scenario,this thesis aims at filling the above-mentioned blanks,and the specific research work is as follows:1.Dataset collection and model selection.Since the lighting conditions,imaging angles,and backgrounds outside the window of the public dataset are relatively monotonous,more samples from real driving and stationary environments are collected as supplements.In order to select a model with high accuracy and good real-time performance,4 classical convolutional neural networks are tested,and finally Mobile Net V2 is selected as the driver distraction recognition model.2.The feasibility of applying federated learning to the scenario of driver distraction recognition is verified.In this thesis,3 sets of distributed data constructed from two datasets are used in experiments,and the experiment results show that even under highly non-iid data settings,the accuracy of the model trained by federated learning does not drop too much compared to that obtained by centralized training method（6%at most）,which proves the feasibility of applying federated learning to the scenario of driver distraction recognition.3.Aiming at the problem that the convergence speed of federated learning drops when the data is non-iid,the fully connected layer gradient constrained federated average method（FCGCFed Avg）is proposed,in this method,the update direction of the fully connected layer of the local model is required to be close to that of the global model of the last round,FCGCFed Avg improves the convergence speed of federated learning by 1.1 to 1.7 times,without increasing any communication overhead,and the increased computing overhead is less than 2%.4.Aiming at the upload communication overhead optimization for federated learning,the federated averaging method based on LCSE（Fed LCSE）is proposed,this method adds loss constrained SE module to the convolution layer of a model.After a node completes local training,only the channel attentions of the SE module and the updated parameters of filters whose channel attentions are equal to 1 are uploaded,which effectively reduces the upload communication overhead by 37% to 40% at the cost of 50% additional computing overhead.5.According to the process of software engineering,the realization of the driver distraction recognition system based on federated learning is briefed from the aspects of requirement analysis,outline design,detailed design,software implementation and testing.