Research On Background Modeling And License Plates Detection Algorithmsin Complex Scenes

The development of computer vision technology laid the foundation for the intelligent analysis of video surveillance. The background of actual surveillance video is usually complex, which contains much interference, noise and complex change. Research of analysis and processing algorithm for video and image in complex scenes is vital in theory and practice. This thesis focuses on the background modeling algorithm and license plate detection algorithm in complex scenes, and obtains the following achievements:First, The machine learning and area-based feature extraction technique is applied to the background modeling, and then one background modeling method is proposed based on the regional characteristics and spatial grid distributed SVM classifier, which combines the single-class SVM classifier and two-class SVM classifier and builds the background model in accordance with the spatial grid distribution of the cascade classifier. The experimental results show that the method in this thesis is robust to the dynamic background and complicated background changes in complex scenes.Second, a background modeling method based on t-distribution is implemented. This method adopts batch approximation algorithm to accelerate the parameter estimation, and enables the model to update online. When predicting, this paper proposes an image acceleration prediction strategy and two related pixel filling methods.Third, for complex scenes, a license plate detection system is designed and implemented which contains four modules of fast locating, license plate detection, character segmentation and vehicles tracking. In the fast positioning module, the following methods are proposed to realize the rapid and efficient locating of license plate:the binarization with self-adaptive threshold value method, the delete of non-license plate region in blocks with two-level thresholds, and the improved connected domain fusion strategy and vertical and horizontal positioning method. In the license plate detection module, we extract the color Haar-like features, and detect the license plate region with SVM classifier. In the character segmentation module, the local binarizition and effective connected domain judgment algorithm are improved. The vehicle tracking algorithm based on the license plate detection is added in video detection to avoid repeat recognition of license plate. The experimental results show that the license plate detection system has good performance in real-time and accuracy.