Research On Traffic Scene Oriented Image Classification

With the flourish of intelligent traffic monitoring technology, it brings the number of traffic surveillance images and videos growing rapidly. It is time-consuming and labor-intensive to analysis all the videos manually. Intelligently fast retrieving and managing traffic images/videos are facing a great challenge. Traffic scene oriented image classification is the ground for traffic image/video intelligently retrieval and management, and it is one of the key technologies to be solved in realizing intelligent monitoring. So the research of traffic scene oriented image classification has theoretical and practical value.The main goal is realizing the traffic scene oriented image classification. This thesis focuses on image feature extraction, image representation and classification. The main research contents of this thesis are as follows:Firstly, Local Binary Pattern based image low-level feature is extracted in thesis. Then support vector machine is adopted to realize the traffic scene oriented image classification. The experimental results do not achieve the desired effect. The main reason is low-level features of images can’t describe image semantic content very well. Visual words can describe image middle-level semantic content. So SIFT is used in this thesis to form the visual words representation of image and support vector machine is adopted to realize the traffic scene oriented image classification. Through comparing experiment results of the two methods, the image classification performance of SIFT based visual words model is better.Secondly, the visual words model ignores the spatial information of image. Spatial pyramid matching model is introduced to represent the images in this thesis. This model which makes use of the image block context in image feature space is the improvement of visual words. This image statement combining with support vector machine is used for traffic scene oriented image classification. Compared with visual words model, the precision is improved significantly.Thirdly. The vector quantization of traditional spatial pyramid model has large quantization errors. The computing complexity of spatial pyramid matching based image classification is relatively high and the run time is too long. In order to solve these problems, locality-constrained linear coding method is introduced to improve the vector quantization coding. This image statement combining with Liblinear classifier is used for traffic scene oriented image classification. This method reduces the computing complexity and running time and improves image classification precision.