DSOL：DSO优化版，加速直接稀疏里程计

需积分: 0 15 浏览量更新于2024-08-03 收藏 3.28MB PDF 举报

"DSOL: AFastDirectSparseOdometryScheme" DSO（直接稀疏里程计）是一种用于实时视觉定位和导航的重要技术，特别是在资源有限的平台如微型无人机上。DSOL（DSO加强版）是DSO的一个改进版本，旨在通过优化算法和实现提升计算效率。以下是DSOL的主要改进点： 1. **逆合成对齐方法**：在帧跟踪阶段，DSOL采用了逆合成对齐，这种方法能更有效地估计图像间的运动，从而提高跟踪精度和速度。 2. **全局窗口跟踪**：不同于DSO仅跟踪最后一个关键帧，DSOL在全局窗口内跟踪新图像，这有助于更好地处理长期的运动和变化，提高了系统的鲁棒性。 3. **双目光度BA公式**：DSOL提出了一种比SDSO（立体DSO）更优的双目光度束调整（Bundle Adjustment, BA）公式。BA是SLAM（Simultaneous Localization And Mapping，同时定位与建图）中的关键步骤，用于优化相机位姿和场景结构，DSOL的新公式在保持精度的同时提高了计算效率。 4. **关键帧管理优化**：DSOL极大地简化了关键帧的创建和删除标准，这降低了内存需求，同时也减少了计算负担，使得系统能更高效地运行。 5. **并行化处理**：为了充分利用现代多核处理器的计算能力，DSOL进行了系统并行化设计，这使得DSO在资源受限的平台上也能实现速度的显著提升，平均提升了5倍。 DSOL的这些改进使得它在处理快速运动场景时表现更佳，尤其适合于资源有限的机器人系统，例如微型无人机。其开源实现可在GitHub上获取，为研究者和开发者提供了实际应用的基础。总结来说，DSOL通过优化核心算法、改进关键帧管理和并行化处理，显著提升了DSO的计算速度和性能，为实时视觉定位提供了更为高效且准确的解决方案。这对于需要在复杂环境中快速定位和导航的机器人系统具有重要意义。

DSOL: A Fast Direct Sparse Odometry Scheme

Chao Qu, Shreyas S. Shivakumar, Ian D. Miller and Camillo J. Taylor

Abstract— In this paper, we describe Direct Sparse Odometry

Lite (DSOL), an improved version of Direct Sparse Odometry

(DSO) [1]. We propose several algorithmic and implementation

enhancements which speed up computation by a signiﬁcant

factor (on average 5x) even on resource constrained plat-

forms. The increase in speed allows us to process images at

higher frame rates, which in turn provides better results on

rapid motions. Our open-source implementation is available at

https://github.com/versatran01/dsol.

I. INTRODUCTION

Localization and mapping are key components of many

robotic systems. In this work we are motivated by the re-

quirements of micro aerial vehicles where payload and power

limit the choice of sensors and the computational capacity

of the system. For these platforms, vision is an attractive

sensing modality since cameras are relatively inexpensive in

terms of mass and power consumption.

We choose to build our work on DSO and Stereo-DSO

(SDSO) as described in [1], [2]. The direct approach to

visual odometry minimizes the photometric error instead

of a geometric one. In their work Engel et al. provide

extensive experiments which show that the direct and sparse

combination offers unique advantages over state-of-the-art

feature-based methods like ORB-SLAM [3] and semi-dense

direct methods such as LSD-SLAM [4] in terms of accuracy

and speed.

The computational advantage is particularly important in

the aerial context that motivates this work, where agile robots

often travel up to 10m/s [5]. We show that the proposed

approach can run at rates of up to 500Hz on common

computational hardware. For direct methods, the improve-

ment in runtime also provides concomitant improvements in

accuracy and robustness [6]: faster processing times allows

for higher frame rates which can then estimate moderate

motions with greater ﬁdelity and aggressive motions with

higher robustness.

For many robotics application, DSO is not suitable since

it is a monocular method which does not have a consis-

tent, metric scale and has delayed initialization and re-

initialization. This poses great safety concerns, especially for

ﬂying robots. For this reason, we design our system to be

directly initializable from stereo and/or depth images. In this

We acknowledge the support of Distributed and Collaborative Intelligent

Systems and Technology Collaborative Research Alliance. Ian Miller ac-

knowledges the support of a NASA Space Technology Research Fellowship.

C. Qu, S. S. Shivakumar, I. Miller and C. J. Taylor are with the GRASP

Laboratory, School of Engineering and Applied Sciences, University

of Pennsylvania {quchao, sshreyas, iandm, cjtaylor}

@seas.upenn.edu

Fig. 1: Results of DSOL on selected sequences from the Tar-

tanAir dataset [7]. Green lines are groundtruth trajectories,

red lines are estimated trajectories.

work, we focus on the stereo version and evaluate our system

against SDSO, which is itself an improvement upon DSO.

The remaining sections describe the proposed system in

more details, highlighting differences compared to DSO and

SDSO, and explaining relevant implementation choices that

lead to improvements. We summarize our changes as follow:

(1) we adopt the inverse compositional alignment approach

in frame tracking, (2) we track the new image w.r.t the entire

window instead of the last keyframe, (3) we propose a better

stereo photometric bundle adjustment formulation compared

to SDSO, (4) we greatly simplify the keyframe creation and

removal criteria from DSO, and (5) we parallelize our system

to effectively utilize all available computational resources.

Together, these changes lead to a simple and lightning-fast

direct sparse odometry system, which we call Direct Sparse

Odometry Lite (DSOL).

II. RELATED WORK

Visual odometry (VO) algorithms can be broadly catego-

rized along the following two axes: direct vs. indirect and

dense vs. sparse. Direct methods recover model parame-

ters directly from images by minimizing photometric error

based on the brightness constancy assumption [8]. This is

in stark contrast to indirect methods, often called feature-

based methods, where correspondences are ﬁrst established

based on some intermediate representations, and the model

parameters are optimized by minimizing reprojection errors.

Dense methods aim to use all information from the image

for better accuracy and robustness at the cost of increased

computation. Sparse methods, on the other hand, recognize

that image data is highly redundant and choose to only

process a selected yet informative subset of the image.

Early VO/V-SLAM systems were mostly sparse and indi-

rect [9], [10], [11], [12]. This was partly due to the com-

putation limit at the time, but also dictated by the needs of

arXiv:2203.08182v1 [cs.RO] 15 Mar 2022

下载后可阅读完整内容，剩余7页未读，立即下载

灵境行者

粉丝: 0
资源: 5

DSOL：DSO优化版，加速直接稀疏里程计

DSOL:DSOL：用Java实现的分布式仿真对象库-开源

一种双目直接法视觉里程计.docx

DSO窗口优化详解：变量与结构解析

DSO203/DS203/DSO quad用户手册：操作与应用指南

双目DSO与IMU融合：柏林自由大学硕士论文

针对跨海大桥的自动化裂缝检测，如何结合大疆M210-RTK无人机与直接稀疏里程计法(DSO)进行裂缝识别与3D建模？

在进行跨海大桥的自动化裂缝检测中，如何利用大疆M210-RTK无人机搭载的高分辨率相机进行图像获取，并通过直接稀疏里程计法(DSO)进行裂缝识别与3D建模？

如何利用大疆M210-RTK无人机和直接稀疏里程计法进行跨海大桥的自动化裂缝检测？

ist的matlab代码-DSO4xx4c:HantekDSO4xx4BC的升级和黑客总结

最新资源