Research And Implementation Of Vehicle Information Extraction Based On Machine Learning

With rapid increasing of car ownership in China, traffic violations such as illegal vehicle modification, vehicle oversizing, are becoming increasingly common. These violations are serious threats to people’s property and lives and need to be detected and corrected right away. In order to identify the overloaded vehicles, this thesis studied and developed a vehicle information detection system to extract 3D dimensions of vehicles. The main task of the system is to identify plate character and detect vehicle dimensions. The main contents of the thesis are as follows:According to the feature of the plate character information, the thesis proposes a license plate recognition algorithm which can adapt the night, backlight, haze and other complex environments. The algorithm is divided into plate location, character segmentation and character recognition. In the section of plate location, license plate is extracted roughly from the background according to the edge and color feature and plate shape. Candidate plates are classified by using the Support Vector Machine (SVM) algorithm of machine learning to get the real plate. In the section of character segmentation, a character segmentation method based on improved connected domain is proposed by analyzing and discussing the traditional methods. In the section of character recognition, the structure and principle of artificial neural network are introduced. Through artificial neural network training relevant character, character recognition is completed.The principle of Binocular Stereo Vision is elaborated in measuring vehicle dimensions. Definition of the image coordinate system, the camera coordinate system and the world coordinate system and their relationship are described. Based on Computer Open Source library (OpenCV) and two-dimensional plane calibration method, camera parameters are calibrated. Principles of Stereo Vision Matching are divided into matching primitive, constraint criteria and stereo matching policy. The thesis selects the local area of the image as image matching primitives to complete stereo matching algorithm, and analyzes the result of stereo matching.The triangulation measurement principle for depth extraction is introduced under the Parallel Binocular Stereo Vision. Using the measurement formula to calculate the distance from camera to each end of the vehicle, vehicle dimensions are calculated. And then, the structure diagram explains the vehicle information extraction system.The thesis applies the measurement and identification technology of computer vision to extract vehicle information. Improved plate location and character segmentation algorithm is proposed. By extracting Plate character information and vehicle dimensions information, vehicle information extraction system is designed.The measurement system of vehicle Information can measure the vehicles automatically without strict background. The system provides an effective technology to identify illegal vehicle modification and vehicle oversizing for relevant departments.