Research On Vehicle Recognition By Integrating Feature Encoding And Convolutional Neural Network

With the development of social economy and urbanization, people's living standards have markedly improved and private cars have become necessary travel tools. The intelligent transportation system research and development have an increasingly important practical value, vehicle recognition, as a key part of intelligent transportation system, has important applications, such as automatic charging system, roads and parking fee, etc.As traditional models based on physical parameters identification algorithms have many disadvantages, such as high cost of installation and maintenance, short service life, large damage to the original pavement, low flexibility, etc. But pattern recognition methods based on video image processing have easy installation and maintenance properties, and does not destroy the traffic road surface with low cost, so this paper mainly research on vehicle recognition on video image processing. Existing methods are mainly based on geometric features whose parameters need to be adjusted when the angle and scene of monitor change. It is not conducive to vehicle recognition system promotion, so this article attempts to use texture features for vehicle recognition. The main work is as follows:In this thesis, we have investigated the latest research about vehicle recognition at home and abroad, and analyzed their advantages and disadvantages. We have improved the original Vi Be's model by changing its update rules to model background, considering the scalability of the pixels spatial distribution and the difference of background sample set distribution, which increase the robustness of the model. We found that it can effectively remove the noise that exist in the foreground. We have compared many traditional image recognition algorithms and convolutional neural network used in vehicle recognition. And,we have presented a new vehicle recognition algorithm based on feature encoding combined with convolutional neural network, which uses CNN can learn higher-level features suitable for classification and recognition that makes model expression capability stronger, and the feature encoding is equal to the pre-training for convolutional neural network. Since it has a certain ability of classification, it can reduce the requirement for convolutional neural network learning to extract features after combined with feature encoding, which speeds up the convolutional neural network convergence and reduces the demand for large training dataset.Finally we have implemented a preliminary vehicle recognition. To implement this system, we first need to extract the foreground video in motion, then extract the prospects and analysis connected component to remove the influence of noise, extract the features of prospects and further encode them into fixed-length vectors which suitable for recognition, finally track on the recognition of certain vehicles models. The final system achieves good results.