The Study On UAV HD Video Image Real-Time Mosaic Algorithm

As a new method of earth observation, the UAV has showed some advantages likeflexibility, high efficiency and convenience etc, that makes it different from the traditionalsatellite system, and it is widely used in land monitoring, resource exploration, disastersurvey etc. However, the images sizes obtained by UAV are usually small, and the numberof video images obtained by UAV is always too large. Thus the UAV jitter can cause someissues such as geometric distortions in images, and the overlap region is irregular graphics.These factors seriously affect the splicing effect and splicing efficiency of video images,and the problems mentioned above will restrict the practical applications of the UAV.Therefore, this topic has important realistic meaning to the research of the UAV HD videoimage real-time mosaic algorithm.An improved registration algorithm considering the shortcomings of the Hu invariantmoments algorithm is proposed. It firstly roughly estimates the overlap region of thematching images according to the information of the UAV flight height, camera angle andvelocity, and extracts edge corners as points of interest in the overlap region. TheEuclidean distance of color histogram around points of interest is calculated as thesimilarity measure to determine the matching bit suspicious according to the scale changeof images. Then Hu invariant moments are used to complete the exact matching of images,and the matching points are tested by RANSAC algorithm. The paper also calculates therotation, scaling, translation and other parameters according to the obtained registrationpoints, and eliminates false matching feature points.The fusion image is preprocessed bythe method of linear energy normalization to balance the gray of the image. Aiming at theproblem that the boundary is often appeared in the fusion images. An improved gradualweight fusion method is also proposed in the paper. This method is different from thetraditional method as the gradual weight which only considers the weight changing in amain direction of the relative position. The weight is calculated by the angle between thecenter-lines of two images with the coordinate axis to adjust the degree. The results show that the algorithm has a good real-time characteristic and can effectively eliminateboundary and improve quality of the fused image. In the end, we summarized what we’vedone in this paper and proposed a vision for the future.