微创手术中内窥镜视觉SLAM方法研究

发布时间：2018-03-23 04:31

本文选题：微创手术　切入点：内窥镜视觉SLAM　出处：《电子科技大学》2017年硕士论文　论文类型：学位论文

【摘要】：内窥镜视觉即时定位与地图创建(Simulation Localization and Mapping,SLAM)是在动态微创手术环境下基于内窥镜实时影像同步完成对软组织环境的三维动态重建和内窥镜自身运动的估计,即获取软组织表面特征在对应场景中的三维空间信息,并通过内窥镜与场景特征的相对位置关系来确定内窥镜在同一空间坐标中的位姿。微创手术特殊复杂的环境,如术中光照不均、软组织流血、诊疗烟雾、软组织形变、软组织表面强边缘图像特征缺失以及湿性软组织表面的高度镜面反射等,对内窥镜视觉SLAM方法的实时性和鲁棒性提出较高要求。针对以上问题,本文开展了微创手术中内窥镜视觉SLAM方法的研究,具体内容如下:1.本文搭建了基于概率估计的内窥镜视觉SLAM框架,主要包括内窥镜及软组织运动建模、内窥镜视觉测量模型以及扩展卡尔曼滤波算法。2.软组织特征测量包括软组织特征提取和特征匹配两个部分。对于软组织的特征提取,本文创新性地应用ORB(Oriented Brief)特征提取方法,保证了提取的软组织特征在微创手术复杂环境中的稳定性;为了获得更好的实时性,本文对内窥镜图像进行了栅格区域划分,根据区域特征分布参数,提取数目稳定且分布均匀的软组织特征。对于软组织的特征匹配,提出软组织特征匹配区域的主动搜索方法,避免全局搜索,从而大大地减小了特征匹配的计算复杂度。最后,将上述软组织特征测量方法代入SLAM框架,并与已有特征测量方法进行对比,实验结果表明,本文提出的软组织特征测量方法具有良好的鲁棒性以及广泛适用性。3.针对扩展卡尔曼滤波存在的缺陷,本文提出基于1点随机抽样一致性(1-point Random Sample Consensus,1-pointRANSAC)的改进扩展卡尔曼滤波算法,删除误匹配的同时补救了有用的软组织特征信息,使得内窥镜定位更精确,并且结合实验对其进行了可行性分析。最后,本文基于达芬奇手术机器人在真实微创手术环境下采集的内窥镜图像数据,对本文所提出的内窥镜视觉SLAM方法进行了验证。实验结果表明,该方法可以有效应对软组织形变对定位和三维重建的动态干扰,软组织特征测量的不确定性椭圆区域能够最终收敛在估计值附近,并且该方法利用较少的软组织特征信息就能实现内窥镜的准确定位,同时能实现内窥镜序列图片7Hz的实时处理速度。
[Abstract]:Simulation Localization and mapping (slam) of endoscope visual real-time location and map creation is to synchronize the 3D dynamic reconstruction of soft tissue environment and the estimation of endoscope motion under dynamic minimally invasive surgery environment based on real-time image of endoscope. That is, the 3D spatial information of soft tissue surface features in the corresponding scene is obtained, and the position and pose of the endoscope in the same space coordinate are determined by the relative position relationship between the endoscope and the scene feature. Such as uneven illumination, soft tissue bleeding, diagnosis and treatment smoke, soft tissue deformation, absence of strong edge image features on soft tissue surface, and high specular reflex on wet soft tissue surface, etc. High requirements for real-time and robustness of endoscope visual SLAM method are put forward. In view of the above problems, the research of endoscope visual SLAM method in minimally invasive surgery is carried out in this paper. The main contents are as follows: 1. In this paper, we set up an endoscope visual SLAM framework based on probability estimation, which mainly includes endoscope and soft tissue motion modeling. Endoscopic vision measurement model and extended Kalman filter algorithm .2. soft tissue feature measurement includes soft tissue feature extraction and feature matching. For soft tissue feature extraction, this paper innovatively applies ORB(Oriented briefing feature extraction method. The stability of the extracted soft tissue features in the complex environment of minimally invasive surgery is guaranteed. In order to obtain better real-time performance, the image of endoscope is divided into raster regions according to the regional feature distribution parameters. For soft tissue feature matching, an active search method for soft tissue feature matching region is proposed to avoid global search, thus greatly reducing the computational complexity of feature matching. The soft tissue feature measurement method mentioned above is put into SLAM frame and compared with the existing feature measurement method. The experimental results show that, The soft tissue feature measurement method proposed in this paper has good robustness and wide applicability. 3. In view of the shortcomings of extended Kalman filter, an improved extended Kalman filter algorithm based on 1-point random sampling consistency and 1-point Random Sample Consensus1-pointRANSAC is proposed. Removing mismatch and repairing useful soft tissue feature information make endoscope localization more accurate. Finally, the feasibility analysis is carried out on the basis of experiments. Based on the image data collected by Leonardo da Vinci surgical robot in a real minimally invasive operation environment, the SLAM method proposed in this paper is verified. The experimental results show that, This method can effectively deal with the dynamic interference of soft tissue deformation on localization and 3D reconstruction, and the uncertain elliptical region of soft tissue feature measurement can finally converge near the estimated value. Using less soft tissue feature information, the method can locate endoscope accurately and realize the real-time processing speed of 7Hz.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：R616;TP391.41

【参考文献】