Rapid Detection And Processing Of Dim And Small Moving Objects

Space surveillance becomes more and more important for apperceiving the situation of space,protecting space resources.Dim and small objects detection is the key component and technology for space surveillance and space situation awareness(SSA).It is an important base for the location of space object in the celestial sphere,determining obit parameters,determining the object size and number of pixels and classification,maintenance of the catalogue of space resident objects(RSO).Dim and small moving objects have a small amount of energy,occupy a few numbers of pixels,and do not have the features of shape and texture.Moreover,its image is similar to the great amount of stars in the image,and it is more easily to form false alarms.Therefore,it is more difficult to detect them than detecting other big objects.There are usually two observing modes,the sidereal stare mode(SSM),and rate track mode(TRM).For SSM mode with long exposure time,an object is a streak(If the exposure time is not very long,the trace of an object is composed of discrete points),while a star is a point source.For TRM mode,a star is a streak,while an object is a point source.This thesis is systematically and deeply focusing on the detection of dim and small moving objects with high efficiency in optical starry background image sequences.(1)Modeling of imaging chain of the astronomical detection systemThe relationship between radiance and star magnitude is built for objects with typical shapes.The point spread function(PSF)and modulation transfer function(MTF)of each links including the optical system,the detector,motion,and the vibration are modeled.The PSF and MTF of the whole astronomical detection system are built.(2)Modeling of optical starry background image with different observing modesThe analysis of noise and background features is important for afterward image preprocessing,while stars and objects modeling is useful for the detection algorithm test and performance evaluation.Firstly,noise is studied and classified according to its probability distribution.Then the features of background,objects and stars in the imaging chain are analyzed and,and the models of objects and stars are built.Finally,all the components in the image are combined and the optical starry background image model is built.The method of generating optical starry background image of TRM mode using images of SSM mode is proposed to test the performance of the detection method.(3)Image preprocessingImage preprocessing contains reducing or eliminating smear effect detected in frame transfer charge coupled device(CCD),which is caused by saturated stars,hot pixels caused by space radiation,and non-uniform background caused by stray light and multi channels of image sensors.Therefore,the appropriate threshold can be used to perform global image segmentation,improve single frame detection and reduce false alarm rate.The hot pixel detection the eliminating method based on local threshold and neighborhoods interpolation,the automatic detection and eliminating method of smear effect,and one dimensional mean iterative background estimating and removal method are proposed to solve the above problems.Besides,the phase transfer function method is proposed to calculate centroids of stars and objects,and the accuracy can reach 0.1211 pixels if the SNR is 3,which is much higher than traditional centroid method.Finally,the image regression method is implemented using phase shift measurement method based on combined transform correlation.(4)Dim and small objects detection in SSM modeMaximum likelihood detection provides a effective mathematic framework,but it has a computational load,and needs the prior information of objects.Optimal projection and maximum projection reduces the computational load,but the performance is sacrificed and can be only used when SNR is median or high.Multistage hypothesis testing(MHT)combines the possible traces of objects as a tree,and prunes the tree at each stage in time to solve the above problems.However,MHT search each pixel in each image,it still has a large computation cost,and it cannot detect discontinuous trajectory of objects.We proposed dim and small objects detection method with time-index image.Firstly,the time index filtering method is used to eliminate stars and noise clearly and quickly.Then,time-index multistage quasi hypothesis testing(TMQHT)method is proposed to detect objects.Search tree is used to detect candidate objects and multistage quasi hypothesis testing is performed to make decision after time index filtering.The proposed method can detect low SNR objects with high probability of detection,low false alarm rate,and a small computational load.Simulation results show that when SNR is larger than or equal to 1.5,the probability of detection is 100%without any false alarm,and the accuracy of location is better than 0.2 pixels.The detection method does not need any prior information of objects,and it can detect those objects with discontinuous trajectory.It needs fewer frames of images than MHT method,and it is easy to implement on hardware platform.(5)Dim and small objects detection in TRM modeThere are a large number of star streaks in the image of TRM mode,therefore,the motion parameters of stars can be computed from the image itself.Then the matched filter can be used to generate the star streak mask.Firstly,the parameter estimation method for star streaks is proposed and derived,which uses both the Fourier and Radon transform.Then,the iterative matched filter method and multi frame summation method is implemented.This method erases stars using the star streak mask generated from iterative matched filtering,and objects can be detected using a large threshold after the SNR is improved through multi frame summation.Simulaion results show that the method can detect objects with SNR equals 1 successfully without any false alarm;when using 5 images as a frame summation set and the SNR is 1.5,the accuracy of detection is 0.1326 pixels.(6)Rapid processing and automatic embedded code generation with multiple DSPs and the fixed point modelingThe rapid processing FPGA+DSP platform is designed based on TI DSP TMS320DM6437 and Xilinx Spartan 3E FPGA XC3S500E.The fixed point model of MATLAB code is optimized and built based on MathWorks tools MATLAB Coder and Fixed-Point Designer.The embedded C code is generated for fixed point DSP TMS320DM6437.