| With the development of transportation technology,intelligent driving has attracted more and more attention.Object detection technology plays a great role in driving scenarios and is an important basis for ensuring the safety evaluation of intelligent driving.The traditional neural network training is based on the central training method.However,the centralized training method may have data privacy issues,such as the leakage of user personal information during the data transmission process.In order to solve the above problems,a federated learning algorithm based on a distributed training method is preferred,which can save user data locally for training,thereby avoiding privacy leakage during data transmission.This thesis studies a distributed training system for deep neural networks based on the federated learning algorithm,which is used for the target detection model training task of driving scenes.The overall federated learning system is divided into two types of devices: server and client,with corresponding network connection modules,model training modules,and aggregation storage modules.On this basis,combined with the model compression technology,the model structure is compressed from the aspects of structure lightening and weight acceleration.Reducing the model volume will speed up the communication efficiency.Experimental tests are conducted for the overall system function and model optimization.The main work of this thesis includes the following two aspects.（1）Use the Pytorch deep learning framework to build a distributed system of federated learning,and combine the database and remote calling functions to realize the tasks of model transmission,model training,and model recording on the terminal.Then,the system based on the selected hardware platform is deployed to obtain the real-time training data of the server and the client and initialize the vehicle detection model in the driving scene,and use this federated learning system to train the model on the terminal to verify the feasibility of the learning system.（2）In order to reduce the size of the network,using model compression technology compress the detection model in different modes to reduce the size of the network,and conduct functional tests and accuracy performance tests on the constructed federated learning system and network model.After two compression experiments of lightweight structure and weight acceleration,the performance changes of the detection model are compared and analyzed,and the experimental results are displayed to verify the practicability and feasibility of different compression methods. |
