Research And Implementation Of Fast Fixed Scene Video Stabilization Algorithms Based On DSP

With the prosperity of video surveillance products,more and more types of surveillance products appear,and are wide-spred into many applications.The cameras of such products are sensitive to outside environmental interferences,such as blowing or camera base shaking,which may lead to jitter of video images.The jitter of videos can cause the observer to suffer from visual fatigue,and may lead to the misjudgment of video contents.Moreover the jitter will seriously degrade subsequent video processing.Therefore,it has become an urgent problem to eliminate the jitter of videos.Video stabilization is one of the important research directions in the field of computer vision.It is essentially a method which takes the geometric transform relation to compensate the irregular motion of image sequences.This thesis mainly focuses on the stabilization of fixed scene videos,whose cameras do not move and monitor a given area.By analyzing the characteristics of video jitter,this thesis investigates the principle and system structure of video stabilization,and some key techniques,including image transform model,global motion estimation,motion compensation and the quality evaluation method of video stabilization.To satisfy practical needs of video stabilization,a new algorithm is proposed to stabilize fixed scene videos and a video stabilization function library based on DSP is developed.Our research work and contributions are described as follows.By simultaneously considering the precision and time cost,a global motion estimation algorithm based on Harris corner detection operator and Pyramid LK optical flow is proposed,and a homography model is applied to describe the image transformation.At the same time,the optimization of feature points,subpixel corner,eliminating interference factors by RANSAC,and the optimization of motion parameters by simplex method are done to improve the accuracy of motion parameter estimation.In motion compensation,an inverse mapping method and bilinear interpolation are used to improve the quality of reconstructed images.By compared with other methods,the proposed algorithm can quickly and effectively solve the problem of fixed scene video jitter.In order to improve the running speed of video stabilization algorithm by making full use of the hardware resources of DSP,5 methods to optimize the code are implemented.The first is to improve software pipelining so that the parallel processing speed can be improved.The second is compiler optimization which can yield efficient codes.The third is the operation conversion from floating point to fixed-point,which can make best use of the computational advantages of DSP.The fourth is using inline functions and software libraries,which can significantly improve the operation efficiency of some functions to the assembly level.The fifth is the design of DMA operation mode,which would make CPU concentrate on computing.For videos with the resolution of 720X480,these optimization works reduce the running time of the video stabilization routine from 2300ms/frame to 40ms/frame and can work in real time.