基于机械扫描声纳图像的AUV视觉同时定位与地图构建研究

Thanks to their characteristics of long range of activities, good maneuverability and concealment, Autonomous Underwater Vehicles (AUVs) have vast application prospect in the industry of both military and civilian. Autonomous localization is the basics and prerequisite for their normal operations. Taking into account the facts that a kind of Mechanically Scanning Sonar (MSS) has the advantages of low cost, small volume and low power consumption, this research focuses on the Simultaneous Localization and Mapping (SLAM) algorithm for AUVs based on MSS. The research will utilize methods of theoretical analysis, computer simulation and real practical experiments to conduct research. In contrast to traditional SLAM with MSS which is based on feature extraction or scan matching, this research creatively converts MSS data into images, and proposes visual graph SLAM framework using the generated sonar images. The main contents include the research on the image generation methods using MSS data, on the front end construction of the pose graph based on direct sonar image alignment and on the loop closure detection method using the MSS sonar images. The research will strive to solve the issues of MSS which are slow sampling speed, noisy data and low information it is able to provide. The research is hoping for providing a novel solution to SLAM for AUVs.

自主水下机器人(AUV)因其具有活动范围大、机动灵活及隐蔽性好等特点，在军事和民用领域具有广泛应用前景。而自主定位是其能够正常作业的基本先决条件。鉴于机械扫描声纳(MSS)具有成本低、体积小、功耗低等优点，本项目拟综合利用理论分析、计算机仿真和实验验证等研究手段研究基于MSS的AUV同时定位与地图构建(SLAM)算法。区别于传统的基于特征提取或者扫描匹配的MSS SLAM算法，本项目创造性地将MSS数据生成图像并根据MSS特点提出基于声纳图像的视觉Graph SLAM算法框架，主要研究内容包括基于MSS声纳图像生成方法的研究, 基于直接声纳图像匹配的Pose Graph前端构建算法的研究和MSS声纳图像SLAM闭环检测方法的研究，着力解决MSS扫描速度慢、数据噪声大和信息量少的问题，从而提高SLAM算法精度和鲁棒性，以期为AUV SLAM的实现提供另一类新思路。

本项目旨在研究利用一种低成本的机械扫描声纳（MSS）实现较高精度的自主水下机器人（AUV）同时定位与地图构建（SLAM）和定位算法。主要研究内容包括：（1）根据AUV运动模型，研究了应用扩展卡尔曼滤波（EKF）融合AUV运动模型、多普勒速度记录仪（DVL）、惯性测量单元（IMU）和深度传感器等信息的航位推算算法；（2）针对MSS存在的扫描速度慢、数据噪声大和信息量少等问题，研究了利用航位推算运动信息和滑动窗口数据结构的扫描生成方法；（3）针对MSS数据特点，研究了对数据关联具有较强鲁棒性的Rao-Blackwellized Particle Filter(RBPF) SLAM算法；（4）针对AUV非线性运动特点，研究了基于MSS数据和已知地图的Monte Carlo定位算法；（5）研究了一种基于多元回归预测方法的视觉定位算法，可用于基于MSS图像的物体定位。经过大量仿真和真实实验数据，以上算法均被证明具有较好的有效性且在SLAM和定位的精度及鲁棒性上优于现有算法，为AUV 的SLAM实现提供了新思路。