【作者】 刘金鑫；

【导师】 张忠民； 邓舸；

【作者基本信息】 哈尔滨工程大学 ， 电子与通信工程（专业学位）， 2021， 硕士

【摘要】 随着我国综合国力的迅速提升,核能源的研究与应用也不断增多。在核工厂蓄水池中会产出大量的核废液,蓄水池中会有一些形状规则的零件散落在池底,打捞零件的工作会通过水下机器人搭载着机械手来完成。水下机器人视觉处理技术的好坏直接影响了水下工作的效率与安全性。因此研究水下光视觉系统具有非常重要的意义。本文针对水下光视觉系统进行研究创新并实现了以下两个功能:1.对水下零件进行三维重建并测量距离及零件尺寸;2.对水下零件进行高精度高实时性目标识别与检测。本文研究内容主要如下:(1)传统水下相机标定方法工作量相对较大且水下图像预处理后的效果并不理想。针上述问题,创新性的将水下折射补偿模型与图像多尺度融合算法相结合提出了一种新型水下图像预处理算法。有效改善了水下图像边缘模糊失真、图像对比度低等问题,算法可靠性较高并且具有普适性。(2)立体视觉技术中的传统立体匹配算法发展进入瓶颈,近年来鲜有创新。针对上述问题,通过对传统立体匹配算法的视差精化步骤进行算法创新,首先深入研究了半全局立体匹配算法及SGBM算法的实现方法。然后提出一种熵率超像素分割一致性检验视差细化算法,算法在Middlebury标准数据集上进行验证,对比传统算法视差精度提升了5.88%。最后采用上述立体匹配算法对水下图像进行实验并得出结果视差图。(3)水下双目立体视觉存在局限性,无法看到目标物体侧面信息,实际应用中往往需要机器人移动位置进行二次拍摄。针对上述问题,提出基于四目视觉的水下光视觉系统,在双目视觉的基础上增加一组相机可以获得目标零件更多的信息以便于进行目标尺寸测算。实验表明,目标零件的测距平均误差为1.7%,八种零件的平均尺寸误差为1.5mm,与同时期国内外现有研究成果对比有一定优势,达到了实验预期结果。(4)水下光视觉系统对水下目标识别与检测要求高精度与高实时性并存。针对上述问题,将在实时性方面有显著优势的YOLOv3算法应用于水下目标识别与检测。实验表明,应用YOLOv3算法对水下单一零件的样本识别率达到93.3%以上,对多种类目标的识别与检测精度达到91.7%,达到了水下光视觉系统对于目标识别功能的预期要求。更多还原

【Abstract】 With the rapid improvement of China’s comprehensive national strength,the research and application of nuclear energy are also increasing at the same time.However,a large amount of waste nuclear liquid can be produced in the reservoir of nuclear factories.Some regular shaped parts will be scattered at the bottom of the reservoir.The work of fishing the scattered parts will be completed by the underwater robot with the manipulator.As the carrier of underwater ecosystem,the visual processing technology of underwater robot directly affects the efficiency and safety of underwater tasks.Therefore,the research of underwater light vision system is of great significance.In terms of the above problems,this paper studies the underwater optical vision system and achieves two new functions: 1.To use the 3D reconstruction technology of underwater parts to measure the distance and part size;2.High precision and real-time target detection of underwater parts.The main contents of this paper are as follows:(1)The workload of traditional underwater camera calibration is immensely heavy,and the results of underwater image preprocessing are not as ideal as possible.Focusing on the problem,combining underwater refraction compensation model with image multi-scale fusion algorithm and proposed a new way of underwater image pre-processing algorithm.The underwater image is compensated by the parameters calibrated by the camera in the air and the refraction rate.And enhanced by multi-scale weight fusion,which improves the underwater image distortion of edge blurred and low image contrast.The algorithm has high reliability and practicability.(2)The development of traditional stereo matching algorithm in stereo vision technology has entered into a period of bottleneck,there has been little innovation in recent years.Focusing on the problem,aiming at the disparity refinement step of traditional stereo matching algorithm.Firstly,deeply study the principle and implementation method of SGBM algorithm.Secondly,a disparity thinning algorithm based on entropy rate super-pixel segmentation consistency test is carried out.The algorithm is validated in Middlebury standard dataset.Compared with the traditional algorithm,the disparity accuracy is improved by 5.88%.Using the stereo matching algorithm mentioned above,the underwater image is tested and obtain the disparity map.(3)The commonly used underwater binocular stereo vision has certain limitations as well,which could hardly see the side information of the target object.In practical applications,it is often necessary to move the position of the robot for secondary photography.Focusing on the problem,a new underwater optical vision system of double-binocular is proposed.By adding an extra group of cameras on the basis of binocular vision can obtain more information of target parts to facilitate the accuracy of measurement of target size.The experimental results show that the average ranging of errors of the target part is 1.7%,and the average size error of the eight kinds of parts is 1.5 mm.Compared with the domestic and foreign dataset in the same period,we have achieved some certain expected results.(4)The underwater optical vision system requires both high precision and high real-time functionality for underwater target identification and detection.Focusing on the problem,this paper proposes to apply YOLOv3 algorithm,which has advantages in the realm of real-time,to underwater target recognition and detection.The experimental results show that the sample identification rate of single underwater part is more than 93.3%,and the recognition and detection accuracy of multiple targets is 91.7%,which meets the expected requirements of underwater optical vision system for target recognition function.更多还原