基于视觉的水上目标测距技术及应用

Technology and Application of Water Target Ranging Based on Vision

双目视觉具有近距离测距精度高、获取的图像内容丰富等优点,是目前无人驾驶领域的研究热点,准确获取双目图像的视差是双目视觉测距的技术关键。然而,水面测距面临目标不规则晃动,增大了双目图像间水上目标的尺度、亮度差异,易造成匹配歧义。为提高立体匹配的准确性,提出一种结合Yolov5进行目标分割和一种改进的SIFT特征匹配方法。通过提取、匹配非线性尺度空间中目标的仿射不变特征点,从而可准确测得双目图像视差,进而得到精准的分割目标距离。实验证明,提出的改进的SIFT算法能够非常准确地得到双目图像的视差值,准确率达到97.1%,匹配时间仅为1.3 s,比原SIFT算法提高了10.4%,匹配时间缩短4 s,比SURF算法提高了11.3%,匹配时间缩短0.5 s。

Binocular vision has the advantages of high short-range ranging accuracy and rich image content.It is currently a research hotspot in the field of unmanned driving.Accurately acquiring the parallax of binocular images is the key technology of bin⁃ocular vision ranging.However,the water surface ranging faces irregular target shaking,which increases the scale and brightness difference of the water target between the binocular images,which is easy to cause ambiguity in matching.In order to improve the ac⁃curacy of stereo matching,a method combining Yolov5 for object segmentation and an improved SIFT feature matching approach are proposed.By extracting and matching the affine invariant feature points of the target in the nonlinear scale space,the image par⁃allax can be accurately measured,and then the accurate target distance can be obtained.Experiments show that the improved model SIFT can obtain the parallax value of binocular image very accurately,the accuracy of the model is 97.1%,the matching time is on⁃ly 1.3 s,which is 10.4%higher than the original SIFT and the matching time is shortened by 4 s,which is 11.3%higher than that of the model SURF,and the matching time is shorted by 0.5 s.