Abstract:

Selection and grading of sea cucumber is an important step for underwater intelligent fishing operations. However, the size of sea cucumber cannot be accurately obtained due to refraction underthewater. So, the underwater binocular calibration method is adopted to calculate parameters of underwater camera and eliminate image distortion caused by the refraction of underwater light. On this basis, an automatic detection and size measurement method for underwater sea cucumber is proposed. Firstly, on the left correction image, the pretrained YOLOv3 detection model for sea cucumber is used to find the bounding boxes of sea cucumbers. In addition, it constructs current depth information of underwater scene by using binocular correction image. Then, a novel GrabCut-RGBD image segmentation method is constructed based on the Gaussian model of fusion color and depth information, and regions of interest with sea cucumber are separated from the background in the bounding box. Finally, by using a convex hull and rotating calipers algorithm, the optimal size measurement points are found on the segmented sea cucumber target image and using triangulation principle to realize automatic measurement of sea cucumber size. The experimental results show that the average error of the proposed method is 1.65% in the range of 0.5~1.5 m, which can fairly realize the automatic detection and size measurement of sea cucumber under the water.

Key words: underwater stereovision, vision measurement, underwater image segmentation, object detection

摘要：

水下捕捞机器人在进行海参捕捞作业时，需要对海参进行精选分级。然而光视觉系统在水中的折射现象严重影响了海参尺寸的精确测量。因此，利用水下双目标定方法，求解水下相机模型参数，消除水下光线折射造成的图像失真，并在此基础上，提出了一种水下海参自动检测与尺寸测量方法。在左目矫正图像上，利用预先训练的YOLOv3海参检测模型，进行海参自动检测和感兴趣区域定位，并利用双目矫正图像构建当前水下场景深度信息。利用融合颜色和场景深度信息的高斯模型，构建新颖的GrabCut-RGBD图像分割方法，在感兴趣区域上分割二维海参目标。利用凸包与旋转卡壳算法，在海参目标图像上寻找最佳尺寸测量点，通过三角测量获取最佳测量点三维坐标，实现海参尺寸的自动测量。实验结果表明，所提方法在0.5～1.5 m范围内平均误差为1.65%，能较好地实现海参尺寸的水下测量。

关键词: 水下立体视觉, 视觉测量, 水下图像分割, 目标检测