[Paper Review] Visual-Inertial Odometry of Aerial Robots

Davide Scaramuzza, Zichao Zhang|arXiv (Cornell University)|Jun 7, 2019

Robotics and Sensor-Based LocalizationEngineering36 references53 citations

TL;DR

This encyclopedia chapter reviews visual-inertial odometry (VIO) for aerial robots, detailing how cameras and IMUs are fused to estimate pose, velocity, and biases, with emphasis on coupled vs. decoupled schemes and smoothing versus filtering approaches.

ABSTRACT

Visual-Inertial odometry (VIO) is the process of estimating the state (pose and velocity) of an agent (e.g., an aerial robot) by using only the input of one or more cameras plus one or more Inertial Measurement Units (IMUs) attached to it. VIO is the only viable alternative to GPS and lidar-based odometry to achieve accurate state estimation. Since both cameras and IMUs are very cheap, these sensor types are ubiquitous in all today's aerial robots.

Motivation & Objective

• Motivate robust state estimation for aerial robots in GPS-denied or challenging environments.
• Explain the complementary roles of cameras and IMUs and why their fusion improves robustness.
• Characterize VIO architectures (loosely coupled, tightly coupled) and estimation paradigms (filtering, fixed-lag smoothing, full smoothing).
• Discuss camera-IMU calibration, synchronization, and online/self-calibration importance.
• Highlight available open-source monocular VIO systems and practical considerations for aerial robotics.

Proposed method

• Describe the VIO state representation and the standard measurement models for camera projections and IMU kinematics.
• Differentiate loosely coupled versus tightly coupled fusion of visual and inertial data."
• Explain the three major VIO paradigms: filtering, fixed-lag smoothing, and full smoothing, including their trade-offs in accuracy and computation.
• Discuss IMU preintegration for real-time smoothing backends to handle high-rate IMU data efficiently.
• Address online calibration of extrinsics and temporal offsets using tools like Kalibr.
• Summarize representative monocular VIO pipelines (MSCKF, OKVIS, ROVIO, VINS-Mono, SVO+MSF/SVO+GTSAM) and their characteristics.

Experimental results

Research questions

• RQ1 What are the main VIO architectures, and how do they trade off accuracy and computation for aerial robots?
• RQ2 How do camera-IMU calibration and synchronization affect VIO performance in UAV applications?
• RQ3 What is the impact of using full smoothing, fixed-lag smoothing, or filtering on VIO accuracy and robustness?
• RQ4 How do open-source monocular VIO systems compare for flying robots in terms of pose estimation quality and real-time performance?

Key findings

• VIO combines camera imagery and IMU data to estimate pose, velocity, and sensor biases, leveraging gravity to resolve scale with a single IMU.
• Tightly coupled fusion generally yields higher accuracy than loosely coupled fusion due to better use of raw measurements.
• Full smoothers provide the highest accuracy but are computationally expensive; fixed-lag smoothers offer a balance; filters are most efficient but may accumulate drift.
• IMU preintegration enables real-time smoothing backends by reusing IMU measurements between updates.
• Camera-IMU spatial and temporal calibration is crucial; Kalibr is a widely used toolbox for both offline and online calibration.
• Several open-source monocular VIO pipelines exist (MSCKF, OKVIS, ROVIO, VINS-Mono, SVO+MSF, SVO+GTSAM), with trade-offs in features, optimization backends, and whether they are stereo/mono.