铁路监控视频的运动估计技术研究

摘要：运动估计是一种估计视频序列中前后帧之间运动位移的技术，块匹配的原理是先将图当前帧划分为若干宏块，然后对当前帧的每一宏块根据一定的匹配准则在相邻帧中搜索当前块的匹配块，得到两者的相对位移，即当前块的运动位移。9460
本文研究的运动估计算法包括全搜索（ Exhaustive Search, ES）、三步搜索法（Three Step Search, TSS）、新三步搜索法（ New Three Step Search, NTSS）和菱形搜索法（ Diamond Search, DS）来估计块的运动位移。全搜索法是对图片中的每个像素进行搜索，得到的最小块误差（Minimum Block Distortion MBD）像素点即为最佳匹配点。三步搜索法是以（0,0）为中心，周围第一步步长为4的九个像素搜索出MBD，第二步长为S/2搜索MBD，直到步长为1得出的MBD即为最佳匹配点。新三步搜索法是搜索中心附近9 9与3 3区域内的17个点，分别搜索找到MBD，若MBD在外围8个点中则进行三步搜索法，若MBD在内部8个点中，则继续搜索以MBD为中心的点得出的MBD即为最佳匹配点。菱形搜索法是对LDSP(大模块)的搜索区域中心以及周围的8个点分别进行匹配计算，当MBD在中心是极为最佳匹配块，否则再以MBD为搜索区域中心以及周围的8个点进行匹配计算直到MBA为中心。
实验是采用峰值信噪比（Peak Signal Noise Ratio, PSNR）和搜索点数来分别评估算法的估计精准度和实时性能。实验结果表明，在这四种算法中，ES算法的搜索点数最大,TSS算法的搜索点数比ES小，但略大于NTSS的搜索点数，DS的搜索点数最小，所以DS实时性能最好。ES算法的PSNR值最大，DS的PSNR值略小于ES，所以ES的估计精度最高。
关键词：运动估计；块匹配；MBD
Motion estimation for digital movie of railway
Abstract: The principle of motion estimation is to pide each frame of video into a plurality of macro blocks and it will not overlap each other.And assuming that all pixels within the macro block are equal displacement, then we use these matching methods to search the same block.The displacement of current block and the reference block is the relative displacement of the displacement movement.
     The principle of the block matching is that the current frame is pided into several macro-blocks, and then each macro block of the current frame according to certain matching criteria to search in the adjacent frame matching block of the current block to obtain a relative
displacement between the two.Then the displacement is motion displacement of the current block.
     Motion estimation algorithm is mainly used the full search (Exhaustive Search, ES), three-step search (Three Step Search, TSS), new three-step search (New Three Step Search, NTSS), and diamond search (Diamond Search, DS ) to estimate the block motion displacement. Full search method is a method witch search each pixel in the picture .The MBD(Minimum Block Distort) is the best match. Three-step search method is based on (0,0) and around 8 pixels, the first step is S=4 to find the MBD, the second length of step is S / 2, until the step size is 1,then the MBD is the best match. New three-step search method is to search the area 9*9 and 3*3 within 17 points to find the MBD ,if the MBD is in the area of 9*9 then with the TSS. If not then search the center of MBD and surrounding 8 points to search out the best match point. Diamond search is search the center and the surrounding 8 points of LDSP to find matching calculation, when the MBD is extremely in the center of the block then the MBA is the best match. If not then search the surrounding eight points until the matching calculation is MBA.
       Experiments use PSNR (Peak Signal Noise Ratio, PSNR) and search points to separately evaluate the algorithm accuracy and real-time performance estimation. Experimental results show that the four algorithms, ES algorithm has maximum search points value, TSS algorithm value is smaller than ES, but slightly larger than NTSS, and DS search points value is minimum, so the best real-time performance is DS.ES algorithm PSNR is maximum value,DS algorithm PSNR value is smaller than ES.Above all, ES is best algorithm to keep accuracy. 铁路监控视频的运动估计技术研究:http://www.751com.cn/zidonghua/lunwen_8188.html