Study On The Method Of Similarity Prior Stitching Of Infrared Images In Low Altitude Remote Sensing

The low altitude remote sensing of UAV(Unmanned Aerial Vehicle)plays an important role in geological survey,precision farming,object tracking and other applications.Image stitching can enable remote sensing to obtain panoramic images with a wide angle of vision and high resolution and improve the efficiency of remote sensing.The acquisition conditions of remote sensing imaging are complex and the images to be spliced will have great differences in the overlapping areas,which pose a great test for the high-precision image stitching.It mainly includes two problems.Remote sensing of UAV mostly adopts sensor intersection stitching technology,and the attitude of UAV will change during flight.Therefore,the actual imaging mode of the imaging system is tilt imaging,and the image obtained from tilt imaging will lead to poor stitching effect.The complexity of remote sensing image will be a big challenge to image registration,the registration must ensure the effect of both overlapping and non-overlapping areas.In this thesis,the core technologies of stitching of infrared images in low altitude remote sensing are studied,including the geometric correction of tilt imaging,the image stitching algorithm based on similarity prior and the construction of the stitching experimental system.The main contents and innovation points are:1.The image correction based on synchronous pose data is studied to improve the quality of the images to be spliced.First,the transformation model from the pixel coordinate system to the local geodetic coordinate system is constructed by using the pose data synchronized with the image data.Then,the image is re-projected to the local geodetic coordinate system for sampling based on the transformation model and the collinear principle,and the distortion of the ladder form caused by tilt imaging is corrected.2.An image stitching algorithm based on similarity prior is studied.Through the optimization method of image transformation based on APAP(As-Projective-AsPossible)and mesh deformation of multi-constraint items(alignment constraint of overlapping region,similarity constraint of non-overlapping region,global similarity constraint of image),the algorithm not only ensures the alignment effect of the overlapped area of the image after stitching,but also reduces the distortion of the nonoverlapped area of the image caused by projection transformation.Among them,SIFT(Scale Invariant Feature Transform)is applied for feature extraction,optimized kd tree(k dimensional tree)algorithm is adopted for coarse matching,sRANSAC(sequence Random Sample Consensus)is used for fine matching,image registration is a method based on similarity prior,and image fusion is a method of weighting in and out.The experimental results show that the mean root-mean-square errors of this method,Autostitch,APAP and AANAP(Adaptive As-Natural-As-Possible)on multiple images are 2.76,8.49,2.96 and 3.00,respectively.3.An infrared image stitching experimental system is constructed,pose data and image data of time synchronization are obtained and a fast image stitching algorithm based on camera’s attitude is studied.In the meantime,the correctness of image correction and the validity of off-line stitching are guaranteed.