Research And Application Of Vehicle Type Recognition Based On Deep Learning

Vehicle recognition is an important part of Intelligent Transportation System(ITS)and has a wide range of applications.Among them,vehicle type recognition is the core link of vehicle identification.Because the vehicle image data has the characteristics of magnanimity,diversity and so on,now the vehicle type recognition task has the higher request to the accuracy and the speed.Traditional vehicle recognition methods are mainly based on artificial design features.Such methods are slow in speed,low in accuracy and difficult to meet current requirements.The deep learning method can autonomously learn the effective features in the image and avoid the artificial design features,so it is of great significance to apply it to the field of vehicle type recognition.Based on the theory of deep learning,this thesis proposes two different vehicle recognition methods based on deep learning.When the labeled data set is sufficient,the spatial pyramid strategy is adopted to solve the problem of vehicle scale change in image.When the labeled data set is insufficient,sparse stack coding network is used to solve the problem of insufficient labeled sample dataset.The main research work is as follows:1.Aiming at the problem that the vehicle scale change in the image leads to the low accuracy of the vehicle recognition method,a vehicle type recognition method based on the combination of spatial pyramid features and depth features is proposed.The core of this method is to use the spatial pyramid method to generate the spatial pyramid features which contain the spatial information of vehicle types.Then the convolution neural network is used to train the spatial pyramid features to generate the depth features containing spatial information.In addition,the trained convolutional neural network is used as feature extractor to extract vehicle image features in training set and train SVM classification model.Experiments on vehicle data set BIT-Vehicle show that the accuracy of vehicle type recognition can be improved by introducing depth features of spatial information.The use of SVM model can effectively alleviate the over-fitting phenomenon.2.Aiming at the low accuracy of the method in chapter 3 of the small sample set,a semi-supervised model recognition method based on the combination of sparse stack coding and convolution is proposed.In this method,sparse stack coding is adopted to learn the data set,and a large number of labels are not needed in training,which avoids the labeling of the data set and improves the degree of automation.At the same time,the idea of sharing weights of convolutional neural network is introduced into feature extraction.The input image is convoluted by using the learned feature dictionary as the convolution core to generate the vehicle feature maps.The dimension is reduced by the maximum pooling method,and the effective features are extracted to represent the vehicle model.By using the Softmax classifier,a small number of label data sets are monitored and fine-tuned.To achieve the purpose of vehicle type recognition.Experiments on vehicle BIT-Vehicle dataset show that this method can further extract higher-level semantic features of vehicles and enhance the expressive ability of the model.Compared with the traditional sparse stack coding method and K-means clustering method,the recognition accuracy in the less labeled sample data set is better than the method proposed in Chapter 3.