Research Of Automobile Type Recognition Based On Embedded System

As an important transmission link in the connection space,the automobile completes transmission of people and things,and plays an increasingly important role in social life.China's auto industry is in a period of vigorous development and rapid increase in the number of cars.And at the same time the city traffic congestion,traffic accidents and car crime occurred.From the middle and late 20 th Century,domestic and foreign began to try to integrate advanced information technology,data communication technology,sensor technology to build Intelligent Transport System(ITS),hoping to fundamentally solve the above series of problems.The efficient,high-precision and fine-grained vehicle type identification system is an important foundation for the const ruction of the Intelligent Transport System(ITS)and its importance is naturally self-evident.On the other hand,because the embedded system itself has a portable,low power consumption,low cost and other characteristics,so that in many aspects(such as field operations)it can do what the PC system can't do.The thesis analyzes and compares the traditional vehicle type identification scheme.In the traditional method,the researcher focuses on the use of what sensors to obtain information about the vehicle type and uses the general classifier(for example,Support Vector Machines,etc.)to roughly discriminate the vehicle type.The method is simple,the theoretical basis is solid and the explanation is strong.However,the use of this method for vehicle type identification requires the engineering staff to have a wealth of experience to be able to construct the hand-craft feature.It is difficult to do the fine-grained categorization of the vehicle.In this thesis,a vehicle type recognition system is constructed based on the Convolution Neural Network method,which combines feature extraction and classification.It is easy to do the fine-grained categorization of the vehicle.Because there is no need to manually extract features,and therefore do not require engineers with a particularly rich experience.In this thesis,the Cars Dataset from Stanford University is used as the basis for constructing the vehicle type recognition system.The effect of different data volume,network type,batch size,dropout value and learning rate on the construction of neural network model is compared.Lastly,the vehicle type recognition system based on the Goog Le Net network model is trained.The relevant Graphical User Interface is designed by Qt tool,and the training model is embedded into the software to realize the vehicle type recognition system based on the embedded system.Due to the above-mentioned neural network model,we need a lot of computation in the process of recognition.So,in this thesis,Jetson TK1 is established as a platform for constructing vehicle type recognition system by comparing several embedded platforms which are more popular today.Finally,the actual test shows that using Jetson TK1 to predict the average of the neural network model is about 228 ms,the performance is excel ent.In summary,it is feasible to construct vehicle type recognition system based on embedded system.