Research On Traffic Scene Image Preprocessing For Driving Assistant

The detection and recognition of the car, lisence plate, traffic sign and so on are determined by the quality of the traffic scene images in intelligent transpotation domain. And the images' quality will also influence the normal function of the driving assistant system. However, due to external noise interference and optical imaging equipment limitation, the acquired images always have some problems such as polluted by the noise and have low resolution in the process of images' collection and transpotaion. Therefore, it is important to research the technologies of image preprocessing under these situations. In order to solve the problems of noisy and low resolution images collected by the vehicle's imaging platform, the image preprocessing technology is presented in this paper. There are three research aspects:Aimed at the problem of polluted images due to external noise interference, a new image denoising method based on wavelet coefficients statistic model is proposed in this paper after introduces the wavelet theory and its statistic characteristics. The two dimensional generalized autoregressive conditional heteroscedasticity(2D-GARCH) model is applied to the wavelet coefficients modeling. This model can make better use of important characteristics of wavelet coefficients such as “sharp peak and heavy tailed” marginal distribution and the dependencies between the coefficients. The maximum likelihood estimation based on fruit fly optimization algorithm is applied to the model's parameters estimation instead of using traditional linear programming and this new algorithm improves the accuracy of the modeling. On that basis the minimum mean square error estimation is applied to estimating the parameters of the original image wavelet coefficients which are not affected by noise to reach the purpose of denoising.Aimed at the problem of low resolution images due to optical imaging instruments limitation, the method of super resolution reconstruction based on sparse representation is applied to the images reconstruction. At first, the application of the sparse represetation theory in the image super resolution domain is introduced in this paper. Then two classical sparse representation algorithms which are based on double dictionaries and structure clustered dictionary respectively are introduced. On that basis a new sparse representation algorithm based on double subdictionaries is proposed which combines with the above two algorithms' characteristics, and this algorithm guarantees the image reconstruction effect basically and reduces the complexity of the algorithm greatly at the same time.The exploitation of image preprocessing system's software interface is completed by using the mixed programming of C# and Mat Lab in the Visual Studio 2008 circumstance. The system software can fulfill the fuctions of the images reading, graying, denoising, super resolution, saving and so on. Then the preprocessing effect is validated by appling vehicle detection algorithm and lisence plate recognition algorithm to the images which are denoised and super resolution reconstructed or not. The result shows that the preprocessing algorithms in this paper are effective.