Research On Multi-camera Panoramic Image Stitching Based Real-time Video Monitoring Technology

It is very critical to achieve videos with wider angle of view in the video surveillancesystem. The implementation techniques of the360degree panorama video stitching based onmultiple cameras hardware platform is addressed in this thesis. On the basis of high qualitymosaic image, the main goal of the algorithm design is to meet the real-time demands of theapplications. Aimed at the application background of panorama stitching in video surveillance,the advantages and constraints of multiple camera panorama stitching system is analyzed inthis thesis. Based on the analysis results, multiple wide angle lenses structure is chosen tocapture the multichannel videos so that a reasonable compromise between the camera field ofview and the number of lenses is achieved. Matching the limit for the cost, the function of thecomplete360degree non-blind area surveillance in the horizontal direction is achieved.Aimed at the real-time performance and the visual effect of the panorama video stitchingalgorithm, the work of this thesis mainly focuses on the following aspects:1、The common CCD sensors and wide-angle lenses are adopted to assemble the cameras,and the multithreading programming technique to capture the multichannel videos is utilized.In this thesis, color correction board based image color correction method is proposed to dealwith the low image quality caused by common surveillance lenses. By the preprocessingusing the proposed method, the image quality is improved. What’s more, an average grayvalue method is adopted to adjust the brightness of the overlapping region in the adjacentimages.2、Based on the classical straight line calibration method, a two-step lens correctionscheme is proposed. According to the proposed straight line method, the lenses are correctedpreliminarily, and then by manually setting and adjusting correction coefficients, the imagesare corrected again. Good correction results can be obtained by using the proposed method.3、According to study the principal of cylindrical projective model, the lens distortioncenter is proposed to substitute the projective center in this thesis. Thus, the problem of thedifferent pixel size of the same objects in the overlapping region which caused by thedistortion of lenses and perspective transform can be successfully solved. An efficient LUTmethod is proposed to execute the lens correlation and cylindrical warping process at once,which is used to improve the performance of the algorithm realization.4、Analyze the principal of the Harris corner based registration and SIFT feature basedregistration. SIFT based registration can get good robustness, but the algorithm complexity is unacceptable; Harris based registration can own low algorithm complexity while the matchingperformance is unsatisfactory on the condition of low image quality which is caused by wideangle lenses. At last, integral image technique based fast normalization cross correlationmethod is chosen to align the images, and the practical experiments demonstrate thefeasibility and effectiveness.5、In addition, a simple multiple resolution linear blending method based on multi-bandblending and linear blending is proposed to optimize the fusion process. On the basis of goodfusion effect, the speed is improved.