Study On Method Of Motion Object Detection And Tracking For Video Satellite

With the continuous development of the application of satellite remote sensing technology,video satellites have become the focus of attention as a new model of earth observation system.Compared with traditional optical satellites,video satellite is mainly characterized by its ability to capture high-definition video in outer space,and obtain dynamic information of moving objects in the observation range through data processing,so as to deeper dynamic application through information extraction and analysis.For the dynamic application requirements of video satellites,there are three basic problems that need to be solved in data processing:first of all,the video satellite adopts the Bayer imaging mode,and each frame of image must be color reconstructed to recover the lost information for subsequent dynamic processing,especially for small-sized targets missing color information;secondly,the imaging trait of the video satellite makes unique complex dynamic changes in the observation scene,which cause seriously false alarms on moving objects detection;thirdly,moving targets of satellite video are generally small-size targets with few pixels,almost no texture features,difficult to describe and robust tracking,and susceptible to be disturbed by proximity similarity.To address the above issues,the basic research method of video satellite dynamic application is deeply studied.Focusing on the breakthrough of video satellite small-size target Bayer color reconstruction,motion detection in satellite video complex dynamic change scene,and inter-frame correlation tracking of weak feature target.The main work and achievements are as follows:(1)To explore the mechanism of high quality Bayer color reconstruction.The imaging model of Bayer pattern and the classical adaptive gradient directed interpolation method are studied.The mechanism to the gradient interpolation method can achieve better results and the importance of gradient judgment to Bayer color reconstruction are analyzed and obtained.The theory of inter-band correlation is studied,and the low-frequency characteristics of inter-band color difference domain are analyzed in frequency domain,and the interpolation reconstruction method based on color difference is studied.The directed interpolation error of Bayer color reconstruction is analyzed,and the essential reason and general rules of interpolation error are revealed.(2)In order to solve the problem which is difficult to give consideration to both effect and efficiency in small size target color reconstruction of satellite video Bayer images,the theory of multi-objective programming and linear solution model are studied.A multi-direction filtered Bayer color reconstruction method based on multi-objective programming model is proposed.For improving the efficiency,an image filtering method based on Fast Fourier Transform is introduced.A set of filters are designed for the proposed method,and the fast color reconstruction of Bayer image is implemented by convolution filtering.The color reconstruction experiment is conduct using real remote sensing image data.Combining the analysis of algorithm complexity and the reconstruction results show that the proposed method can give consideration to both effect and efficiency,the PSNR of green band is greater than 40.It is beneficial to subsequent processing of moving object detection and tracking.(3)To explore the special scene change caused by video satellite imaging,the geometric model of video satellite area array detector is deduced.The method of construction and solution of rational function model is studied.The compensation model of stereo pair orientation based on rational function model is deduced,and the relationship between spatial transformation is established.Through the imaging mode of video satellite,we analyze the motion model of scene change from the angle of external orientation elements,reveals the uniqueness of the change of satellite video scene.Analyze the influence of the uniqueness on satellite video motion detection.(4)Aiming at the problem of pseudo motion false alarm caused by the unique scene change of satellite video,a motion detection method of satellite video based on decision tree weak classification is proposed.The differentiation motion of the scene is divided into two parts:global and regional.For the global motion,the median flow optical flow method is studied,and the block matching of the median flow is used to construct the inter-frame orientation correction paramete.The inter-frame motion model based on rational function model is established,and the method of inter-frame motion compensation and compensation parameter solution are studied.For local motion,the decision tree theory and model construction method are studied,and the decision tree model of motion detection is established by analyzing the information gain of target characteristics.The classical ViBe algorithm is improved,and the mechanism of updating differentiation background model based on dynamic updating factor is proposed,which can effectively solve the pseudo motion false alarm problem.The error detection removal rate is greater than 90%.(5)The classical Bayesian theorem and Bayesian estimation are studied,the basic methods and principles using naive Bayesian classification for target tracking are studied.The conditional probability likelihood measurement method based on target attributes is studied.We study and summarize the main difficulties of target tracking under the condition of satellite video.Analyze and reveal the essential reasons for the poor tracking effect by classical methods for satellite video point target.(6)Aiming at the point target tracking problem under satellite video condition,the robust feature description method based on satellite video target feature is studied.In this paper,we propose a satellite video target tracking method based on Bayesian classification framework which combines target dynamic and template features.The probability likelihood method combining target dynamic and template features is studied and a conditional probability classifier model for target tracking is established.The tracking accuracy is greater than 90%under(?)threshold.The research results of the above three basic problem can provide reference for the basic processing of video satellite data,and can also provide the necessary foundation and support for the subsequent application.