RP-VIO: Robust Plane-based Visual-Inertial Odometry for Dynamic Environments

TL;DR

RP-VIO is a monocular visual-inertial odometry system that exploits static planes like walls and ground to enhance robustness and accuracy in highly dynamic environments, supported by a new synthetic dataset.

Contribution

The paper introduces RP-VIO, a novel plane-based VINS that improves robustness without relying on semantic classification, and provides a new synthetic dataset for evaluation.

Findings

RP-VIO outperforms state-of-the-art VINS in dynamic environments.

The synthetic dataset enables better evaluation of VINS robustness.

Plane geometry significantly enhances odometry accuracy in dynamic scenes.

Abstract

Modern visual-inertial navigation systems (VINS) are faced with a critical challenge in real-world deployment: they need to operate reliably and robustly in highly dynamic environments. Current best solutions merely filter dynamic objects as outliers based on the semantics of the object category. Such an approach does not scale as it requires semantic classifiers to encompass all possibly-moving object classes; this is hard to define, let alone deploy. On the other hand, many real-world environments exhibit strong structural regularities in the form of planes such as walls and ground surfaces, which are also crucially static. We present RP-VIO, a monocular visual-inertial odometry system that leverages the simple geometry of these planes for improved robustness and accuracy in challenging dynamic environments. Since existing datasets have a limited number of dynamic elements, we also present a highly-dynamic, photorealistic synthetic dataset for a more effective evaluation of the capabilities of modern VINS systems. We evaluate our approach on this dataset, and three diverse sequences from standard datasets including two real-world dynamic sequences and show a significant improvement in robustness and accuracy over a state-of-the-art monocular visual-inertial odometry system. We also show in simulation an improvement over a simple dynamic-features masking approach. Our code and dataset are publicly available.