全方位视觉大图像并行解算方法研究

全景视觉的大图像中,单DSP难以在规定的时间内完成对图像的处理要求,尤其在连续视频图像上的处理存在延迟.基于此设计了3种方法：图像均分4份;图像按内外环分割;奇偶场分割;并用于多DSP进行处理分割后的图像,再进行合并,以达到实时性的要求.图像均分4份处理速度要优于其他2种,但这种方法运用DSP的数量也要多于其他2种,另外该方法在相对分辨率较低的情况时,竖直方向会存在毛刺问题.按内外环分割,并行处理的时间要按其中像素点较多,处理时间较长的环来计算,而且拼接处接缝处的细化也要延长处理时间,需要注意r的选取.而按奇偶场分割,是将图像的像素点均分,有效解决了毛刺和缝隙问题.仿真采用C＋＋语言在VC＋＋环境下进行.

Research on the parallel reverse solving methods for the omnidirectional large images

This paper shows some methods to solve the omni directional large images processing speed which is so slowly in single DSP （Digital Signal Processor） chip processing. The single DSP chip can not finish the process in time, especially in the processing of continual video images, which mostly are delayed. Based on this problem, the paper supplied three methods to solute them： average the image into four parts; cut the image into inside circle and outside circle; and cut the image for odd and even lines. These methods were designed for more DSP chips to process large images and combined them to meet real time requirement. The first method processes faster than other two, while the number of DSPs is more than the other two. In addition, in the case of relatively low resolution, there are some glitches in the vertical direction. In the second method, the parallel processing time should be calculated according to the circle images which have more pixels and need longer processing time, further more, the refinement of seam in the joint delays the processing time, so we must care the choice of r. The third method averages images and effectively solves the problems of glitches and seams. The simulation was done with C ＋ ＋ in the environment of VC ＋ ＋.