Research On Vehicle Brand Classification Methods Of Multi-Feature Fusion

The report "Development Strategy of Intelligent Transportation during 2012-2020" issued by the Chinese Ministry of Transportation, pointed that the perception and monitoring of traffic information are the key technologies to realize urban intelligent traffic development. As an integral part of ITS, vehicle type has become the basis of data processing and analyses of public transport service, operation and security. Currently vehicle type recognition is maturely used in road electronic toll and parking management, which can’t meet specific needs like identifying a decked car. With the technology development of computer vision and pattern recognition, using image processing to achieve specific vehicle brand identification has become urgently to be solved in vehicle management and maintenance.In this paper, some related researches were done on vehicle brand recognition through feature fusion of HD images captured by cameras. The specific contents are as follows:Firstly, a vehicle brand recognition method based on the texture feature was researched. According to the relative position of the license plate and vehicle face, an area was cropped as the studying object and a car-face database was established including 30 common brands for the training and testing. Then it compared accuracies of local energy shape histogram (LESH), Local Binary Pattern histogram (LBPH), Histogram of oriented gradients (HOG) and other descriptive characteristics classified by Support Vector Machines (SVMs) and k-Nearest Neighbor (kNN). The experimental results show the HOG feature and SVM identification method is superior to the others, whose overall recognition rate is 92.40%.Secondly, the paper studied the vehicle brand classification method based on the sparse feature. It made some analyses and comparisons between the traditional non-dictionary sparse decomposition and sparse coding method based on K-SVD dictionary learning. Through the principle of minimizing the reconstruction error, vehicle brands were classified. Experimental results show that the K-SVD dictionary learning method of classification based on sparse representation has a higher accuracy than the traditional one. And with the increase of the number of training samples and atoms, the recognition rate increases gradually, whose highest recognition rate is 89.07% and the average identification time is less than 0.015 seconds.Finally, an efficient vehicle brand identification method based on multi-feature fusion through sparse representation. It analyzed the combined strategy of the image feature fusion and then the HOG feature was extracted on the first level. Due to the non-negative HOG feature, a non-negative sparse coding method was drawn to form a secondary abstract fusion of HOG feature through sparse representation. With a minimum reconstructed error determining principle, the overall recognition rate is up to 96.16%, and the method is highly robust to light reflection, weak irradiation or local occlusion. Therefore, the proposed frame based on multi-feature fusion for vehicle brand recognition model, which is a high-precision and robust method having better value for application and popularization.