Research On Real-time Image Stabilization Method Of UAV Video Based On Scale Space

As an emerging aviation detection platform,UAVs are characterized by low cost,flexibility and high efficiency,and can easily collect video data.It has been widely used in the fields of investigation and monitoring,panoramic shooting,rescue search and natural resource monitoring,and has gradually become a hot research topic in various countries.During the flight of the drone,the jitter caused by the body or the natural environment will seriously affect the image capturing effect and reduce the value of the captured resources.And many shooting resources are time-sensitive,and it is necessary to obtain effective aerial video in time to facilitate subsequent research and use.Therefore,fast real-time digital image stabilization technology has a good application prospect.The process of digital image stabilization mainly includes motion estimation,motion filtering and motion compensation.In this paper,based on three-dimensional geometric model,image matching and iterative filtering,a stable and efficient real-time digital image stabilization method is proposed.In the traditional image stabilization principle and implementation method,combined with the aerial photography characteristics of the drone and the actual shooting environment,the image is stabilized by the previous frame of the current frame of the video,and the iteration is calculated smoothly from the first frame.The stable effect combines the improved AGAST algorithm with the improved adaptive Kalman filter to achieve real-time image stabilization of the aerial video of the drone.The main work of the text is:(1)Image features are detected using the improved AG AST algorithm.The AG AST algorithm is based on the static structural features of the image and is extremely fast.Since AGAST itself does not have scale invariance,it is not well suited for feature detection of drone video images.This paper will combine Gaussian scale pyramids,create scale space by Gaussian operator LoG,and calculate AGAST features on each layer to make scale invariance,rotation without deformation and illumination invariance.Then use FREAK binary coding to make feature descriptions.Compared with traditional feature description methods,binary description mode calculation and feature matching are faster.(2)Eliminate mismatched points,achieve efficient and accurate motion vector estimation,and facilitate subsequent iterative calculation.The aerial video range of the drone is large,the viewing angle is wide,and the amount of information contained in each frame is large.The Hamming distance matching method based on the binary descriptor will generate a mismatching point.In order to ensure the matching accuracy and calculation speed,and to separate the foreground area and the background area,the RANSAC algorithm is used to eliminate the mismatched points and multi-dimensional similar matching points,and the interference of image noise is eliminated to improve the subsequent processing speed and accuracy.(3)Obtain stable image stabilization results by adaptive filtering.The global motion vector calculated by any method contains the motion vector and the jitter vector.The jitter vector can be regarded as a kind of noise interference,and the jitter vector is filtered by Kalman filter to obtain the vector of the active motion of the drone.Kalman filtering is a time-domain filter based on motion continuity,which is calculated in a time-series iteration.It does not require historical information recording,and is especially suitable for real-time processing.However,in practical applications,the estimated state may be biased.The accumulation of such errors may cause severe divergence and the image stabilization effect of the entire video.Through the time-varying noise estimator,the noise variance matrix is calculated and corrected in real time to achieve the adaptive effect and effectively avoid the divergence phenomenon.Experiments show that compared with the traditional feature detection and motion filtering methods,the proposed algorithm has a significant effect.It can effectively and effectively eliminate video jitter and robustness in continuous video sequence processing,and can be effectively applied to drone aerial video.The image stabilization processing and fast calculation speed can be effectively applied to the real-time processing system.In the background of the vast application of drones,such as earthquake rescue,weather warning,forest fire prevention and control and vegetation observation,the image stabilization algorithm has a wide application market.