移动设备辅助的机器人高性能DSO系统研究与实现

Development of Mobile-Device Assisted DSO (MADSO) for Robust and High-Precision Robot Mapping

樊梦严

11749192

硕士

计算机技术领域工程

郝祁

2020-06-02

2020-07-20

哈尔滨工业大学

深圳

直接稀疏里程计 (Direct Sparse Odometry，DSO) 是直接法中一种相对新颖且精度较高的稀疏视觉里程计。DSO对基于最小化光度误差的概率模型中涉及到的参数，进行一致的联合优化，从而达到较快的处理速度与精度。作为一种直接 SLAM 技术，DSO 在基于官方数据集的跟踪和半稠密建图中有较好的表现，但是在实际应用中，DSO需要提高对光照变化、光度条件以及模糊运动的处理能力。此外，DSO作为视觉里程计，在建图过程中并未进行回环检测，不能有效消除累计误差，其鲁棒性与精度仍有较大提升空间。基于以上问题分析，本文研究并实现了移动设备辅助的机器人高性能DSO系统（Mobile-Device Assisted Direct Sparse Odometry, MADSO)，该系统有效提高了机器人应用DSO建图的精度与鲁棒性。MADSO系统具备的创新点或优势有：为有效缓解DSO算法在实际应用过程中，初始化与位姿估计时漂移等问题，MADSO借助移动设备跟踪机器人Tag标志，为机器人位姿估计初始化与建图灵活地提供观测信息，有效提高了系统初始化与位姿估计的精度。为优化机器人在不同应用场景下位姿估计的误差，MADSO分析了移动设备追踪机器人常见的三种情况，设计并实现了移动设备在不同追踪效果下的解决方案。同时，为降低移动设备追踪丢失机器人与重追到踪机器人间产生的误差，本文提出并实现了平滑轨迹算法，该算法可有效提升机器人轨迹估计的精度。 本文的主要贡献有:分析DSO算法优势与弊端，调研并对比提升DSO算法鲁棒性与精度的相关研究，总结现有方法存在的技术挑战。设计实验并验证DSO算法在实际应用中存在的尺度漂移等问题。统一移动设备与机器人坐标系，计算移动设备在系统坐标系下位姿，根据移动设备跟踪机器人得到的观测位姿与移动设备位姿，对机器人精确定位，辅助机器人完成初始化过程。借助动作捕捉系统采集机器人及移动设备真实运动轨迹，搭建实验验证平台，对比验证MADSO系统位姿估计的精度与鲁棒性。本文的实验验证部分根据动作捕捉系统采集的数据，与DSO、LDSO以及ORB-SLAM算法对比验证位姿估计性能。实验表明，在光照和光度条件变化的情况下，本文所提出的MADSO方法对机器人轨迹估计误差和三维建图精度有了较大的提高。

Direct Sparse Odometry(DSO) is a novel and high precision sparse odometry method of the direct SLAM. DSO uses probabilistic models and joint model parameter optimization methods to minimize photometric errors with faster processing speeds and higher accuracy. DSO has a good performance in pose localization and semi-dense mapping based on datasets. However, in real-world applications, DSO needs to improve its high sensitivity to illumination and photometric conditions as well as motion blurring.Moreover, DSO has no loop closing module, so it cannot effectively reduce the accumulated errors, and hence its robustness and accuracy have a large room for improvement.Based on the above analyses, this paper proposes a direct synchronous localization and mapping method assisted by using a mobile device to improve the accuracy and robustness of DSO, i.e., MADSO. The noveltys or advantages of MADSO are as follows.This work can provide observation information for robot initialization and mapping by tracking robot with mobile device to effectively alleviate the problem of DSO scale drift in real-world applications.MADSO developped solutions for three typical situations when the mobile device tracks the robot for different applications to optimize the error of pose estimation.Meanwhile，we developped a trajectory smoothing algorithm to improve the odometry’s accuracy when the mobile device loses the track of the robot.The major contributions of this work include:Analyzing the advantages and disadvantages of DSO, comparing the possible methods to improve the accuracy and the robustness of DSO, and performing experiments to reveal the problems of DSO in real-world applications, such as scale drift and illumination sensitivity.Developing a framework to estimate the pose of a tagged robot using a mobile device, whose pose is estimated with a lightweight SLAM scheme, which can help initialize the DSO based robot mapping.Setting up an experiment platform to verify the performance of MADSO with the help of a motion capture system.

DSO 机器人建图 移动设备 非线性优化

DSO robot mapping mobile devices nonlinear optimization

中文

联合培养

南方科技大学

樊梦严. 移动设备辅助的机器人高性能DSO系统研究与实现[D]. 深圳. 哈尔滨工业大学,2020.