Research On 360-Degree Panoramic Stitching Algorithm In Assisted Driving

Assisted driving is one of the main research directions in the field of intelligent connected vehicle.Its 360-degree panoramic stitching system provides drivers with environmental information around the car to ensure safty in complex traffic environments.It is the most basic part of assisted driving.In order to solve the problem that the panoramic view of the current panoramic stitching system is small and the stitching quality is poor,the following researches are conducted in the dissertation:Firstly,the preprocessing and feature detection algorithms in panoramic stitching are.Aiming at the problem of aberration and luminance inconsistency in fisheye image,the omnidirectional camera calibration method and the luminance correction method based on local gain compensation are selected as the preprocessing algorithm in this dissertation through experimental comparison.Then,three feature detection algorithms,SIFT,SURF and ORB,are studied respectively.In Mikolajczyk data set,the quantitative,real-time,repeatability and accuracy of each algorithm are compared and analyzed.SURF is selected as the improved object of the algorithm in this dissertation.Secondly,feature detection and matching algorithm in panoramic stitching are proposed.LNMS-SURF algorithm is proposed to solve the problem of non-uniform spatial distribution of SURF feature points,which resulted in redundancy of feature points and incomplete expression of image information.Aiming at the problem of poor real-time performance of traditional feature matching algorithm and small number of correct matching pairs,MMF-GMS feature matching algorithm is proposed.First,KD tree search is adopted,the ratio test,the threshold of matching difference and the bidirectional cross-check multi-matching strategy are fused to carry out the coarse matching of feature points,and then GMS algorithm is used to achieve the purpose of matching pair purification.Experiments on Mikolajczyk data set show that,compared with other matching algorithms,the correct matching rate of this algorithm is improved by 19.6% on average,and the real-time performance is improved by about 3 times.Thirdly,the registration algorithm in the panoramic stitching is improved.In view of the traditional matching algorithms low alignment accuracy and the Non-overlapping area distortion problems,an LLT-GST registration algorithm for the fusion of local homographic matrix and global similarity transformation matrix is proposed.Firstly,Moving DLT is used to solve the local homographic matrix of each image mesh to improve the alignment quality of overlapping areas.Then the local homographic matrix of the overlapped region is linearized and extrapolated to the non-overlapped region.The global similarity transformation is used in the non-overlapping region to further reduce the distortion.Finally,the linearized local homographic matrix is gradually changed to the global similarity transformation matrix.Experiments on multiple data sets showed that the RMSE of the proposed algorithm is 7.149%,superior to the algorithms of AutoStitch,APAP and SPHP.Finally,the fusion algorithm based on the improved stitching seams is analyzed and the experimental platform of the assisted driving panoramic stitching system is built.In order to eliminate the ambiguity caused by the double image and parallax caused by moving objects,adopted the best stitching seams algorithm based on graph cutting to obtain the stitching seams.And then compared the advantages and disadvantages of the feather fusion and multi-resolution fusion algorithm.Finally,the experimental platform of the driving-assisted panoramic stitching system is built to verify the panoramic stitching algorithm.The results show that the algorithm in this dissertation can generate panoramic images with a wide field of vision and high quality,which is of high practical value.