基于全景视觉的无人船水面障碍物检测方法

Detection Method of Water-Surface Obstacles for Unmanned Ships Based on Panoramic Vision

无人船航行时水面障碍物检测因视角不足,导致漏检或误检,同时为满足无人船安全正常作业的需求,提出基于全景视觉的无人船水面障碍物目标检测方法。与传统的单目和双目视觉相比,全景视觉具有水平方向大视场监控的优点。基于多目全景视觉系统获得待拼接图像,在加速稳健特征(SURF)算法的基础上进行图像配准,引入k维树来构建数据索引,实现搜索空间级分类并进行快速匹配。通过M估计样本一致算法对匹配点进行优化,剔除误匹配点。对于图像融合中重叠区域出现的拼接缝隙或重影问题,设计一种基于圆弧函数的加权融合算法。提出改进的水面障碍物目标检测模型DS-YOLOv5s,将拼接好的全景图像作为训练好的模型作为输入,从而检测目标障碍物。实验结果表明,改进后的SURF算法与SURF算法相比特征点的匹配正确率提高11.47个百分点,在匹配时间上比SURF、RANSAC算法缩短5.83 s,DS-YOLOv5s模型的mAP@0.5达到95.7%,检测速度为51帧/s,符合实时目标检测标准。

During the navigation of an unmanned ship,an inadequate perspective may cause obstacles on the water-surface to be missed.To enable the safe and normal operation of unmanned ships,this study proposes a panoramic vision-based method for the detection of water-surface obstacles.Panoramic vision is employed because it has the advantage of horizontal large-field monitoring,in contrast to traditional monocular and binocular vision.In the proposed approach,a multi-camera panoramic vision system acquires an image to be stitched.The Speeded-Up Robust Feature(SURF)algorithm then performs image registration.A k-dimensional tree constructs a data index,facilitating search-space level classification and fast matching.The M-estimation-based sample consensus algorithm optimizes the matching points and eliminates the mismatched points.A specifically designed arc function-based weighted fusion algorithm stitches the gaps and overcomes ghosting in the overlapping areas during image fusion.Finally,this study proposes an improved water-surface obstacle target detection model DS-YOLOv5s.This model takes the stitched panoramic image as input to detect target obstacles.In experiment,the improved SURF algorithm improved the feature-point matching accuracy by 11.47 percentage points compared to the SURF algorithm,and shorten the matching time by 5.83 seconds compared to the SURF,RANSAC algorithm.The DS-YOLOv5s model mAP@0.5 reached 95.7%,with a detection speed of 51 frames/s,conforming to real-time object detection standards.