Research On Key Techonologies In Intelligent Transportion Surveillance Systems Based On Finite Ridgelet Transform And Compressed Sensing Theory

Intelligent transportation surveillance system gains increasing attention with the fast development of mobile transportation. It targets at transportation information collection, providing assistance in decision making, store and management of related information in computer vision manner. How to capture useful information is undebatable the most important task in such surveillance systems. As compared with traditional detection and surveillance methods such as ground loops, computer vision based transportation surveillance system is able to provide more comprehensive information in a faster way. However, the captured transportation images and videos are easily affected by illumination variation, occlusion and environmental disturbances. Their varying image quality pose great challenge for subsequent image processing tasks. For the problems of how to develop effective representation methods, recognition algorithms of traffic signs and license characters plate and encoders which can make full use of intrinsic characteristics of transportation images and videos, they still exist. Based on finite ridgelet transform and compressed sensing theory, this dissertation focuses on algorithm improvement and their applications in transportation image denoising, license plate character recognition and transportation surveillance video compression.Aimed at sparse representation of digital images, design of effective representation and analysis methods has long been hot topic in the signal processing society. Multi-scale geometrical analysis methods, which is developed relatively independently from the subjects of harmonic analysis, computer vision, and statistical analysis, etc. has provided several novel transform methods, such as ridgelet transform, curvelet transform, contourlet transform, bandelet transform and wedgelet transform and has been intensively discussed recently. Wavelet is sensitive to horizontal, vertical and diagonal directions only due to its build in definition, and thus falls behind in effective representation of geometrical structures in digital images. To overcome the weakness of wavelet, Multiscale geometrical analysis theory provides multi-resolution and multi-directional representation of images. As one of the multiscale geometrical analysis tools, ridgelet transform aims at better representation of linear singularities of images. Its basic idea is firstly use Radon transform to map linear singularities in spatial domain to point singularities which are later handled by the wavelet transform. With the consideration of abundant linear singularities in roads, license plates and traffic signs, finite ridgelet transform which is defined in the discrete domain is adopted as the basic image representation tool in this dissertation. Excellent coefficient compactness provides possibility for effective removal of noise from images. Meanwhile, features extracted in the finite ridgelet transform domain are expected to obtain satisfying result in license plate character recognition application.Compressed sensing theory emerged in recent decades aims at sparse representation of signals by means of constructing over-complete dictionaries. It utilizes the intrinsic sparsity characteristic of signals to bypass the traditional sampling process based on Nyquist Sampling Theorem. Instead, the sampling is achieved by means of firstly design measurement matrix according to specific properties of signals followed by an optimization process for solution of a l1norm minimization problem. Researchers develop successful applications in human face recognition in occlusion environments by firstly constructing dictionary using features from all training samples, and then measuring the linear expansion coefficient sparsity of testing sample with respect to the training dictionary. Experimental findings reveal that compressed sensing based classifier relaxes dimension reduction requirements of feature selection which is a quite appealing property. Application in license plate character recognition is studied in this dissertation.With the observation of fixed scenarios and self-similarity property of images and videos in intelligent transportation surveillance system, a block-based compression algorithm is proposed to exploit the spatial redundancies for improved compression performance. Motion estimation and motion compensation techniques are utilized to encode still images and I frames in videos.The main work of this dissertation includes:(1) Algorithm improvement of finite ridgelet transform.Based on analysis of weakness of two dimensional separable wavelet transform in high-dimensional signal representation, basic theories of continuous ridgelet transform is reviewed. Finite Radon transform and transform selection scheme in finite Radon domain are discussed in detail. Relationship with other transforms is also presented. An energy based adaptive finite ridgelet transform algorithm is then proposed, together with the proposal of related columnwise threshold for image denoising application. Performance comparison with other finite ridgelet transforms is carried out to show its validity.(2) Transportation image denoising based on finite ridgelet transform.Based on characteristics analysis of transportatio images, the previous proposed algorithm of energy-based adaptive finite ridgelet transform (EFRIT) with columnwise threshold is applied to traffic image denoising. Performance comparisons of different transform and thresholding strategies are discussed based on abundant experiments. Considering thelarge dynamic range of aspect ratios of transportation images in real applications, a block-overlapping-based denoising algorithm is proposed, which is quite computational efficient. Simulation results on traffic sign database verify its effectiveness.(3) Research on license plate character recognition algorithm based on finite ridgelet features and compressed sensing classifiers.Based on review of features and classifiers for license plate character recognition algorithms, finite ridgelet coefficients of the input license plate character images in lexicographical order are proposed to be extracted as feature vector. Its robustness in presence of translation, noise and occlusion disturbances is examined carefully. In terms of classifier design, features of all training samples are utilized to form overcomplete dictionaries, with respect to which sparse representation of input test samples is solved via minimization of l1norm. Classification results are given based on sparsity measurement of such representation vectors. Simulation results on license plate character image database reveals that the compressed sensing classifier with lexicographical finite ridgelet coefficients as feature vector is robust to noise as well as occlusion.(4) Transportation surveillance video compression based on blocking matching.A blocking matching based compression algorithm is proposed, where motion estimation and motion compensation techniques which are traditionally used in video coding are adopted for better exploitation of spatial redundancy in still images in blockwise layer. Experimental results on natural images, scalable video coding and transportation surveillance video coding demonstrate performance improvement in compression efficiency. Furthermore, the proposed block matching module can be flexible generalized to any video encoder.