A Research And Implementation Of Airborne Video Stitching

Video Stitching technology merges two or more frames to one single frame by feature extraction,image align,perspective transformation and image fusion.In the case of UAV capturing,video stitching can convert multi-channel input videos to single channel output,thus reducing the capacity of transmitting data and the complexity of image processing system.Traditional video stitching algorithms usually have problems like: high computation complexity,low robustness and frame shake.To solve these problems and stitch better panorama,several improvements have done in this paper:(1)In this paper,image registration based on feature matching and multi-scale optical flow tracking is combined to enhance image aligning work.Key frame is aligned by feature matching algorithm,and optical flow tracking algorithm for other frames’ registration.This hybrid image registration method can make full use of the information of frame stream and reduce the time consuming.(2)Traditional P-KLT algorithm has good performance in small optical flow calculation.But it frequently failed to track the parameters in motion blur frames.The robustness is low in these cases.This paper is devoted to optical transmit between P-KLT layers.Multi-scale homography revising algorithm has been added to the original P-KLT to reduce the error between different layers.The homography revising process is based on consistency of local optical flow.Compared with traditional algorithm,the improved algorithm has lower layer errors and higher robustness.(3)The stitching results often suffer unwanted shake which is bad for watching.A video stabilization algorithm is applied in the stitching part.The stitching result is stabilized by parameter estimation,parameter filtering and motion compensation.The experiment results have higher stability and continuity.(4)The video stabilization algorithm in this paper uses homography model which is estimated in frame stitching algorithm to avoid duplicate computation.Compared with similar transform and affine transform,homography transform has higher freedom to describe 3D geometry between consecutive frames.Video stabilization algorithm based on homography model has more stable outputs.An image distortion correction algorithm is proposed to correct the homography distortion.(5)There are undefined image zones in video stabilization results.Two methods is used to do image compensation: image compensation based on sweeping window and local area compensation algorithm.These two methods can both increase image content and enhance viewing effect.