Research On UAV Image Stitching Technologies Based On GPU

Nowadays,the technology related to UAV is developing rapidly,especially the technology that use UAV equipped with high-definition camera to obtain image data.But a single UAV image cannot display the complete information of an area.To get the UAV image which have higher resolution and larger view field,this thesis researches the technology of UAV image stitching,and uses the GPU to accelerate the stitching process.In addition,taking UAV aerial photography in a foggy environment will greatly affect the imaging quality of the image,and the problem of blurred image feature information may occur,which may result in the failure of normal feature point extraction afterwards.This thesis considers that using image defogging in the pre-processing,and studies the method of image defogging.The specific research contents are summarized as follows:(1)In foggy weather,the feature information of UAV image will be blurred,and even the feature points cannot be extracted normally.In order to better extract the feature points from foggy UAV images,this thesis proposes an adaptive image defogging method based on the dark channel.Using the atmospheric scattering model,to divide the image into sky and non-sky area.To solve the transmittance,the color attenuation prior is used in the sky area,while the dark channel prior is used in the non-sky area.Finally using the weighted guided filtering to smooth image.The experimental results show that,no matter which the subjective or objective evaluation is,the method proposed in this thesis has better effect after defogging than other classic methods.It can preserve as much detail of the image as possible which is the foundation of the feature point extration.(2)The traditional SURF algorithm has a very long time overhead beacause it only uses CPU for calculation.In addition,its matching accuracy of feature points is low.In order to improve the process calculation speed of the image stitching process,and ensure the accuracy of the feature matching,after analyzing the traditional SURF algorithm,this thesis proposes the FSURF algorithm.Combined the GPU's highly parallel architecture,the FSURF algorithm uses the GPU to accelerate the calculations of the highly parallel computing processes in the SURF algorithm,such as integral image calculation,feature point extraction,and feature point description.The process of feature point matching uses the two-way fast approximate nearest neighbor algorithm to roughly match,then uses the PROSAC method to remove the false matching of feature pairs.The experimental results show that the running speed of the FSURF method is faster and the matching accuracy is higher than that of the traditional SURF method.(3)Due to the different exposures of the two images to be stitched,the image fused by the traditional best suture method has a sudden change of color in the overlapping area,there will be an obvious splicing crack.In order to eliminate the sudden change of color,this thesis improves the best stitching algorithm.After finding the best suture,it uses the weighted fusion method to smooth both sides of the best suture,and ensures that the transition of the stitched image is more natural.