Optimization And Integration For Vehicle-mounted Video Image Processing Algorithm

Vehicle-mounted video monitoring system is a product which combines thedisciplines of computer vision, pattern recognition, signal processing, artificialintelligence, machine learning. It usually consists of the vehicle terminal, wirelessnetwork sub-system, and the central management platform. This thesis focuses on thekey technologies in the vehicle terminal, that is, the collection of video imagesstabilization, the enhancement and noise reduction of images, the detection andtracking of the moving vehicle, the encoding and compression of video images.The video monitoring system is often mounted inside the vehicle, any shaking orbumping will cause vibration in the video image collected, which will constitutedifficulties for the driver’s observation and the follow-up image processing. Thecapacity of SD card for video image storage is usually very limited. What’s more,adverse surroundings inside and outside the vehicle, such as dim light, hazy and rainyday, darkness at night, will all make the video images collected unclear and full ofnoise. Therefore, the intelligent video monitoring system has becom e a researchhotspot. Much attention has been paid to the recognition of the target location,appearance and the moving track from a complex dynamic background for users’examination and retrieval and for provision of element information on higher-levelvideo comprehension. And that is the task of moving target detection and tracking.This thesis proposes a series of improved algorithm for the image stabilization,image enhancement, noise reduction, target detection and tracking, and video imagecompression of image collected by vehicle-mounted video monitoring system. Themajor research work of this thesis are as follows:(1) IntroductionThis part introduces the function, technological difficulties, system compositionand the developing tendency of the vehicle-mounted video monitoring system. Theresearch status and developing tendency of the related algorithm of image processingare outlined. The research background, significance, emphasis, technical route, thesisorganization are presented.(2) Stabilization Algorithm of the Vehicle-Mounted Video ImageA digital image stabilization algorithm based on sub-block image matching incentral zone is proposed. Firstly the basic principle of traditional sub-block image matching, algorithm flow, and the features of local moving vector are expounded. Theinfluence of changing video background on the local moving vector is analyzed.According the features of the local moving vector in different image zone, thematching sub-blocks are extracted from the central image zone. By doing so, theinfluence of changing video background on the sub-block matching accuracy can bereduced. Simulation experiments showed that this method, compared with thetradition matching method, can significantly reduce the influence of changi ng videobackground on image stabilization.(3) Algorithm of Video Image EnhancementAccording to human vision characteristics, a multiscale Retine enhancementalgorithm for the degraded image caused by bad weather such as rain, fog, and so onis proposed. Simulation experiments show that this algorithm, compared with themethods of homomorphic filtering, HE, MSR, can improve the contrast and sharpnessof the image, better restore the image color effectively, and better control the noise.Another algorithm for enhancement of image at night is proposed as well. With thatalgorithm, the image collected at night is transformed from RGB color space to HIScolor space, the global adaptive adjustment is applied to the luminance component,the trilateral filter is then used to collect the illumination component while the borderis maintained, the luminance component is compressed so as to enhance the contrastbetween the local image and the details. The restore of image color is achieved bysaturation component enhancement. The function of global luminance adaptiveadjustment can effectively stretch the dynamic range between the dark part and thehighlight part of the image. The illumination component of light image is collected bytrilateral filter. The disadvantage of edge blur and passivation during the process ofimage enhancement is overcome. Compared with the method such as Gammacompression, gradient domain, and bilateral filter and so on, the method gets rid ofthe halo in the highlight part, increase the details and restore the color.(4) Recognition Algorithm of Moving Target in Video ImageFirst, Gaussian motion model of moving vehicles in a dynamic scene wasproposed. And the differences between moving vehicles and motion vector on thedynamic background was studied.And the two Gaussian motion model presented in thispaper,the models of moving vehicles and motion vector on the dynamic backgroundwere set up. Adopting Bayesian framework, the motion pixel in the scene was classifiedinto the moving vehicles or dynamic background. The experimental results show thatthe Gaussian model can effectively detect moving vehicles in a variety of dynamic scenes. The robustness is good. In solving the problem of dynamic background noise,the proposed algorithm was better than background subtraction and Gaussian mixturemodel.(5) Coding algorithm of Video Image CompressionMP coding method based on block partition is proposed for low coding efficiencyof the previous algorithm and the generated un-embedding code stream. MPdecomposition based on redundant dictionary is discussed from the dictionaryconstruction and sparse decomposition. With this algorithm, MP atom energy anddistribution features are made full use of, the atom parameters and atom locationparameters are effectively organized and encoded, and thus the coding efficiency isimproved and the embedded code stream is generated. From the low to the middlecode rate, the objective distortion performance equivalent to JPEG2000and SPIHT isobtained and the better subjective quality is achieved as well, which shows theadvantage of sparse decomposition. In addition, while achieving good codingperformance, this algorithm maintains the advantages of MP image coding, andprovides greater flexibility of quality and resolution ratio than the traditional methodsdo, which makes it more applicable to video monitoring system.(6) ConclusionFinally, the main research content and achievements are summarized, theshortcomings are indicated, and the future research direction is stated.