Core Concepts
Tightly integrating deep dense bundle adjustment (DBA) with multi-sensor information enables real-time dense mapping in large-scale environments.
Abstract
The content discusses the integration of deep dense visual bundle adjustment with multiple sensors for large-scale localization and mapping. It covers the framework, system implementation, experiments on different datasets, and real-time performance analysis.
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021
DBA-Fusion integrates trainable deep dense DBA with multi-sensor information.
Recurrent optical flow and DBA are performed among sequential images.
The system supports flexible integration of multiple sensors for large-scale applications.
Extensive tests validate superior localization performance enabling real-time dense mapping.
INTRODUCTION
Visual simultaneous localization and mapping (VS-LAM) is crucial in VR/AR and robotics applications.
Deep learning has significantly advanced VSLAM accuracy and robustness.
Incorporating multiple sensors like IMU, GNSS, and WSS enhances system stability and scale observability.
SYSTEM IMPLEMENTATION
Recurrent optical flow computes dense pixel association for image pairs in a co-visibility graph.
Integrating DBA into a generic factor graph tightly fuses geometric information with multi-sensor data.
Multi-sensor factor graph solves pose estimation efficiently using GTSAM optimization.
EXPERIMENTS
TUM-VI Dataset
DBA-Fusion shows superior performance compared to other VIO algorithms on challenging sequences.
Online mapping results demonstrate improved consistency over DROID-SLAM.
KITTI-360 Dataset
DBA-VIO outperforms other monocular schemes in relative pose errors evaluation.
Optical flow tracking benefits from IMU aiding in achieving better convergence.
Self-Made Urban Dataset
GNSS RTK scheme achieves driftless positioning but faces errors during occlusion.
DBA-Fusion with wheel speed or GNSS integration provides stable decimeter-level position estimation.
CONCLUSION
DBA-Fusion tightly integrates deep dense visual bundle adjustment with multiple sensors for real-time localization and dense mapping. Future work aims to extend the system to dynamic scenarios and neural map representations.
Stats
The proposed method shows dramatically better translation and attitude estimation than the visual-only DROID-SLAM, verifying the contribution of IMU integration to maintain low-drifting, metric-scale pose estimation.