Research On Vehicle Type Recognition Method Based On Improved Faster R-CNN

With the rapid improvement of China's economic level,the continuous development of the vehicle industry,and the continuous improvement of residents' living standards,the number of cars also keep improving,which poses huge challenges to China's traffic management system.In order to further improve the efficiency of traffic management,the establishment of intelligent traffic management system(ITS)is essential,and the identification of different types of vehicles according to the needs of daily traffic management is an important part of the intelligent traffic management system.However,the current vehicle identification methods such as vehicle model identification and vehicle category identification can't satisfy the needs of daily traffic management.The thesis has conducted research on different types of vehicle identification,and mainly divided vehicles into nine categories according to traffic management needs,namely police vehicles,ambulances,fire fighting vehicles,private cars,bus,trucks,construction vehicles,two-wheelers vehicles and other kinds of vehicles,and carry out identification tasks.Based on the characteristics of different types of vehicle images,the thesis starts with the structure of Faster R-CNN.Improved the recognition method and the accuracy of recognition.The main work of the thesis is as follows:1)Studying different CNN structure principles,then compare the advantages and disadvantages of them,also need to compare the mainstream methods in the field of target detection at this stage,such as YOLO,Fast R-CNN,SPP-Net,SSD,Faster R-CNN,so that more clearly understand the advantages of Faster R-CNN.2)Combining the tasks of different types of vehicle recognition.The thesis improves some of the structures on the basis of the original Faster R-CNN network structure to better solve the practical application problems of different types of vehicle recognition and realize its application value.According to the difference of some different types of vehicles in local areas such as vehicle painting and ceiling lights,the thesis chose to introduce a convolution method of hollow convolution instead of traditional convolution to increase the size of the receptive field.Add an empty space pyramid pooling(ASPP)module to improve the network response to feature information at different scales.3)Combined with the characteristics of the data set,the thesis introduces an attention mechanism to improve the network.Based on local features,a method of using cascaded network to modify the recognition results is proposed to reduce the false detection rate between vehicles of the same model but different types,and a comparative analysis of the overall experimental results.4)Based on the method proposed in the thesis,different types of vehicle identification systems are designed and implemented,confirmed the effectiveness of the method.