Study On The Key Technology Of Intelligent Transportation System Based On The Theory Of Image Recognition

Along with the increment of global motor vehicle, intelligent transportation system(ITS) is increasingly becoming the important mean to solve the problem of modern transportation. The multidisciplinary ITS involves image processing, intelligent recognition, machine vision, based on the theory of image recognition of intelligent parking technology, license plate recognition and vehicle recognition technology are the key technology widely applied in intelligent transportation system. The paper will research by the premise of image recognition technology in those three fields:1. We propose the intelligent parking system algorithm based on the camera calibration model. This algorithm combines the pinhole imaging principle in the camera calibration model with parking kinematics model of real-time image processing method, which solves the problem of the steering angle determination and trajectory calculation for intelligent parking. To improve the accuracy of the parking trajectory calculation, we make the further research on he curve of the planning parking trajectory and camera calibration model. First of all, we analysis the variables and the space collision points in trajectory constraint algorithm, after that we establish intelligent parking constraint equations and propose multi-stage arc advance and retreat parking trajectory algorithm. Then based on the full of studying zhang zhengyou two-dimensional plane calibration method, to simply the solving process and improve the calibration precision and robustness, we introduce tangential distortion coefficients in the original distortion model and propose the initial value optimization. We conduct a large amount of experiments about collecting frame rate and track precision with custom experimental vehicle, the experimental data indicated that the algorithm for track precision is higher, the system runs stably, screen showing is fluent, good real-time performance.2. We propose improved algorithm based on fractal dimension and characteristics of hidden markov license recognition system. The license recognition technology will be divided into three steps to study.(1)In the process of license image preprocessing and binarization, we proposed an improved adaptive multilevel median filter algorithm to deal with the noise of image. Then we put forward an improved image binarization based on the differential box fractal dimension of gray image binarization method.(2)In the process of license location and character segmentation correction, first based on adding two directions templates in the traditional Sobel algorithm of image edge detection and redistributing weight, we proposed an improved algorithm based on Sobel operator license positions precisely; then proposed character correction algorithm based on Radon transform through to improve panning and zooming of Radon. Finally, we proposed license single-character segmentation improved algorithm based on the vertical projection through add appropriate parameters to vertical projection algorithm. After optimization of the above-mentioned algorithms, the license image detection accuracy is improved, the edge details become more delicate, continuous and accurately positioning, simultaneously maintaining the original character topology and reducing the distortion of characters.(3)In the process of character recognition, we present a license character recognition algorithm based on hidden Markov model, in which discrete cosine transform is employed to achieve the conversion from light intensity data to frequency data, and the difference matrix is computed from the viewpoints of conditional attribute sets and individual subsets to get the threshold value. Finally we combine multiple classifiers to finalize the license character recognition. By comparing with other state-of-the-arts, it can observe the method has high accuracy rate for the license recognition.3. We propose research on combing BP neural network and HOG feature extraction for improving vehicle identification algorithm. Firstly, we will study separately on the vehicle identification and other identification models, then taken together both of them as the vehicle identification basis.(1)In the car logo recognition algorithm, we propose an improved SIFT operator combing with BP network. The method makes use of license and vehicle–logo relative position to locate the vehicle-logo location, and adopt non-fixed ring valued and increase the weight coefficient to solve the problem of largest calculation and time complexity due to high dimensionality traditional SIFT feature descriptor, finally for extracting the logo SIFT feature descriptor to identify by BP neural network algorithm.(2)In the vehicle identification algorithm, we propose the improved HOG features combine with SVM classifier. The extracting HOG features according to the determined outline feature of models. The efficient and accurate vehicle recognition achieved by SVM classifier to train. Experiment demonstrates that the improved algorithm has high recognition rate and has strong robustness for light, partial occlusion, and noise.The above several algorithms in ITS have a wide application prospect, it could able to make appropriate recommendations for safe driving by the intelligent parking technology application of intelligent vehicle navigation. License recognition and vehicle identification technology can be applied in the field of traffic management and traffic monitoring, thus can come true vehicle detection, vehicle tracking, traffic flow parameters detection, incident detection and acquire amount of traffic data. It is great significance for improving traffic capacity, reducing traffic accidents, properly regulating the traffic flow of distributing road network.