With the rapid development of intelligent traffic monitoring technology, the number of traffic surveillance images and videos is showing explosive growth, the demand of image classification technology is becoming increasingly urgent. Traffic scene oriented image classification techniques is used to automatically classify the traffic scene image into different categories with computer technology. It’s also one of the key technologies of intelligent traffic monitoring with a very good prospect.The main objective of this thesis studies the traffic scene-oriented image classification techniques. We focuses on traffic scene oriented image feature extraction, image representation and classification. The main contents are as follows:Firstly, the BOF model is used to extract the SIFT feature of image to form a visual vocabulary expression in this thesis, and then the SVM classifier is adopted to traffic scene oriented image classification, but the image classification performs not very well.Secondly, due to the BOF model ignores the spatial information of the image, this thesis introduces the spatial pyramid matching model to improve the characteristics of the bag of features model. The SPM model makes use of the image block context in image feature space. Compared with BOF model, the final results of this model significantly increased.Finally, the error of vector quantization in the traditional BOF model and SPM model is large, and the computational complexity of image classification is relatively high, it also requires a long time. To improve the vector quantization coding method and SIFT descriptors invariance, this thesis introduces a local soft assignment coding and Color SIFT descriptors, and combined liblinear classifier in traffic scene image classification, it effectively improves the image classification accuracy rate, and reduces computational complexity and running time. |