Vehicle Detecting Algorithm Research Based On Computer Vision And RVM

Vehicle detection technology is an important part of Intelligent Transportation Systems(ITS), and is one of the core technologies as well. The vehicle detection technology based on computer vision has played a vital role in vehicle detection applications, meanwhile, the vehicle classification detection plays an essential role in promoting the rapid development of the ITS. Therefor, three relevant issues are researched in this thesis, as follows.An algorithm of image nonlinear denoising and enhancement in the Nonsubsampled Contourlet Transform (NSCT) domain by combining with Generalized Cross Validation(GCV) principle is proposed. Firstly, the corrupted image is transformed into NSCT domain, then the optimal denoising threshold of the high-frequency coefficients is estimated with GCV principle in each direction at corresponding level in NSCT domain, respectively. Then soft-threshold shrinkage is implemented to the high-frequency coefficients of the noisy image at corresponding level. The threshold value estimated in previous step is used to remove the noise in the image. Finally, the processed coefficients are enhanced by a nonlinear function to enhance the image edges. Experimental results show that the proposed algorithm can effectively remove noise in the case of unknown noise variance, and enhance the image edges, and achieve the desired effect from both visual effect and objective evaluation standards. The comprehensive performance of the proposed algorithm is better than the methods based on Wavelet Transform(WT) and Contourlet Transform(CT).An algorithm of lane detection based on Hough transform combined with Inverse Perspective Mapping(IPM) is proposed. Firstly, the inner parameters (mean the intrinsic parameters such as visual angles, resolution ratio, etc.) and outer parameters (mean the fixed pose of camera such as the pitch angles of optical axis, the yaw angle, the height apart from road surface, etc.) of camera are initialized. Then the actual road image is transformed by IPM to convert the radial lanes which originally collapsed at one point into paralell ones for estimating the forane traffic conditions and improving driving safety. Finally, the lanes are detected by Hough transform and marked to give some useful informations as the position reference values of vehicle, the driving speed, the degree of congestion. In addition, background subtraction is adopted to extract vehicle objective area, and the Time Temporal Median Filter(TTMF) is adopted during the processing of background extracting in this thesis. In this method, median filtering based on time domain is implemented to the corresponding position pixels of number of odd frames before current frame in the video sequences, so that the background is effectively extracted. The method of shadow detection and elimination based on computer vision is adopted. In this method, the objective image contained shadows is transformed into HSV color space, then the shadows are detected and removed according to their visual properties. Experimental results show that the proposed lane detection algorithm has a nice accuracy, especially detecting terminal lane, and it is easily implemented in practice. In addition, for the background extracking based on TTMF is actually ideal and less time consuming, the vehicle objective region is perfectly detected. From above all, the proposed algorithm has good comprehensive performances.An algorithm of vehicle classification detection with various characteristics based on Relevance Vector Machine(RVM) is proposed. Firstly, morphology operation is implemented to the vehicle objective binary image obtained in previous chapters to more approximate to the actual shape of the objective, then the maximum connection domain of the objective is calculated as the vehicle objective area in following processing progresses. Then various classification characteristics of the objective area and its corresponding gray image are extracted, such as region descriptors, invariant moments, Discrete Cosine Tranform(DCT) descriptors, fractal dimension, and textures; then the characteristics data is normalized, and the data sample set is divided into train samples and test samples. Finally, the RVM is trained to establish the classification detection model, then the obtained model at previous step is tested by test samples. Experimental results show that the classification detection effect of the proposed algorithm is very well, and classification accuracy of test samples is nice. The test speed is obviously pretty ideal compared with those methods based on SVM and BP neural network, so this feature is very useful to improve system real-time performance.