Thesis - Table of Contents

[surberj]

OUTDATED: for current version see comment below !

Abstract

Introduction

• Aeroworks
• Aeroworks Requirements regarding one-agent Localization (Task 4.1)
• VI-Slam / VI-Odometry:
  • Goal of SLAM
  • Why Camera and IMU ?
  • state of the art: Filtering vs Keyframe based Slam
• VI-Sensor

  General Problem Formulation

• State Representation (Robot state + Landmarks)
• State Propagation with IMU
• Landmark representation: FEATURES
  • Common Features
  • ROVIO: Multilevel Patch Features
  • OKVIS: Multiscale SSE-optimized Harris corner detector combined with BRISK

    ROVIO approach

• state: pos, quat, vel, bias, feature-positions wrt robot
• EKF: IMU propagation and Camera update
• Feature selection
• Most important params

  OKVIS approach

• state for optimization: non marginalized keyframe poses, non-keyframe poses, speed and biases and keypoints (landmarks)
• IMU Propagation and Nonlinear batch optimization based on all state
• Keypoint and Keyframe selection
• Marginalization

  Challenges

• Drifting / loop closure detection
• Outdoors
• Scalability - Computational Complexity
• Time Synchronization
• Sensor Noise
• Calibration

  Results

• Evaluation Metrics (translational absolute trajectory error, translational relative pose error, computational complexity: timings)
• Results on different datasets (indoors / outdoors)
• ROVIO insights (Limits)
  • IMU timestamp bias
  • Number of Features

    Conclusion / Outline

[surberj]

Infos, which have to been provided for any dataset:

• source of ground truth (vicon, dgps, simulation)
• how was the data acquired (handheld, quad, bicycle, ...)
• senors
• duration
• travelled distance
• motion characteristics (maximum translational velocities, maximum rotational rates)
• environment (indoors, outdoors, illumination, lots of "landmarks", ...)

[surberj]

Abstract

Introduction

• Why (single agent) localization ?
• Why camera and IMU ?
• Filtering and keyframe based approaches, Introduce State-of-the-Art and OKVIS and ROVIO
• Terminology: Odometry vs SLAM
• Drift issue --> Motivate why I did this analyses
• Time-synchronization issue --> Motivate why I did this analysis
• Main contribution:
  • A direct comparison between OKVIS and ROVIO
  • Demonstrating Global and Local Accuracies of the 2 algorithms
  • Deeper Insights into the drifting issues of ROVIO
  • Demonstrating ROVIO running on a non hardware-wise time synchronized sensor system

Methods

ROVIO

• Basic Principle
• Fast Corner Detector
• Multilevel Direct Intensity Patch "Features"
• Small number of Landmarks (circa 25)

OKVIS

• Basic Principle
• Multiscale SSE-optimized Harris corner detector
• BRISK Features
• Marginalization
• Optimization over a few hundreds to thousands of Landmarks

Evaluation Metrics

• Generell Notation: Trajectory with position and attitude
• Alignment in time and space
• Global accuracy - absolute translation and orientation error, yaw-error, world-z error
• Local accuracy - relative translation and orientation error
• Processing time per frame as metric for computational complexity

Results, Discussion

• IJRR dataset - ROVIO and OKVIS results
  • Dataset
  • Global accuracy
  • OKVIS better
  • ROVIO drifting in yaw 45 deg / 1000 m - 10 times higher than OKVIS
  • roll and pitch bounded < 1 degree - 2-3 times higher than OKVIS
  • world-x and world-y driven by yaw error,
  • world-z error 1m / 1000m - 3 times higher than OKVIS
  • Why the difference ?
  • What to do about it?
    • Global Optimization,
    • Include additional Sensor Information - especially magnetometer, eventually GPS
  • Local accuracy
  • OKVIS slightly better but similar
  • relative translation error 2 cm / 1 m
  • relative orientation error 0.6 deg / 1 m
  • Interpretation: Is this meaningful for a task like dense reconstruction?
  • Computational Complexity
• EuRoC dataset - ROVIO results
  • Global accuracy,
  • Local accuracy,
  • Processing time and accuracy depending on number of features
  • Discussion: Which differences and Why to IJRR set
• Non hardware-wise time synchronized sensor system
  • What means hardware-wise time synchronized ?
  • results on analysis with artificially added timestamp bias
  • results on my datasets
  • ROVIO running with hardware-wise non-timesynchronized sensor system,
  • fixing the camera-imu extrinsics can increase performance

Conclusion, Outline

• Global Accuracy:
  • OKVIS better / drift issue within ROVIO, mainly yaw-drift
  • Future Work: Global Optimization
  • Include magnetometer to tackle yaw-drift
• Local Accuracy
  • OKVIS slightly better, both similar and reasonably good for a task like dense reconstruction
• Towards non hardware-wise time synchronized sensor systems
  • ROVIO demonstrated to run with non hardware-wise time-synchronized sensor system
  • If high dynamics are expected, fixing the camera-imu extrinsics could enhance ROVIO performance
  • Future Work: Run ROVIO on embedded system with non hardware-timesynchronized sensor system

ev Acknowledgment

• Michael Blösch, Bharat Tak, Franziska Ryser, Marco Karrer

References

[surberj]

Report is handed in, https://github.com/surberj/Vision/tree/master/thesis/Report