Fine Vehicle Type Recognition Based On Deep Learning

The purpose of fine vehicle type recognition is to classify the vehicles with small appearance differences,which is widely used in the fields of statistical traffic flow,highway automatic toll collection and driverless.However,there are still the following problems in the existing fine vehicle type recognition methods: On the one hand,most of the methods stay in the research of a single front angle of bayonet monitoring,and this kind of methods fail to take into account the uneven distribution of image features,which leads to the poor recognition accuracy;On the other hand,the research on multi angle is not extensive enough.The main reason is that the background of vehicles in multi angle is complex and diverse,and the visual characteristics are not stable,which brings difficulties to the recognition task.Based on this,the main research work of this paper is as follows:(1)According to the uneven feature distribution of the front angle vehicle image under the traffic checkpoint monitoring,a fine vehicle type recognition algorithm based on the front image is proposed.Firstly,based on convolutional neural network,two branch networks with different depths and a fusion network are designed.Among them,the shallow upper branch network is used to extract non vehicle surface features,the deep lower branch network is used to extract vehicle surface features,and the fusion network is used to carry out multi branch and multi scale fusion of the features extracted from the two regions,so as to further abstract and extract features with significant discrimination.The experimental results on the front angle vehicle data set captured by the bayonet show that the recognition effect of this method is better and has practical application value.(2)Based on the analysis of the difficulties of multi angle fine vehicle type recognition,such as complex background and unstable visual features,a vehicle fine recognition algorithm based on multi angle feature fusion is proposed.Firstly,SSD vehicle detection framework based on improved resnet-50 is used to detect and locate multi angle target vehicles,After that,the trained resnet-50 network is reused to extract vehicle features from different angles,and the AUC weighted fusion strategy is used to fuse the extracted features,Finally,the softmax loss measure learning method with adjustable class spacing is used for recognition.Theexperimental results on the multi angle vehicle data set show that the method in this paper has achieved good recognition results.