OV$^{2}$SLAM : A Fully Online and Versatile Visual SLAM for Real-Time Applications (2102.04060v1)

Abstract: Many applications of Visual SLAM, such as augmented reality, virtual reality, robotics or autonomous driving, require versatile, robust and precise solutions, most often with real-time capability. In this work, we describe OV$^{2}$SLAM, a fully online algorithm, handling both monocular and stereo camera setups, various map scales and frame-rates ranging from a few Hertz up to several hundreds. It combines numerous recent contributions in visual localization within an efficient multi-threaded architecture. Extensive comparisons with competing algorithms shows the state-of-the-art accuracy and real-time performance of the resulting algorithm. For the benefit of the community, we release the source code: \url{https://github.com/ov2slam/ov2slam}.

• The paper introduces a multi-threaded, fully online visual SLAM algorithm that integrates an online bag-of-words loop closure to enhance mapping accuracy.
• It supports both monocular and stereo inputs and processes high-frequency data up to 400 Hz, outperforming methods like ORB-SLAM on benchmark datasets.
• Its adaptive design ensures robust localization in diverse settings, making it ideal for applications in autonomous navigation, augmented reality, and urban mapping.

OV$^{2}$SLAM: A Comprehensive Overview

The paper presents OV$^{2}$SLAM, a Visual Simultaneous Localization and Mapping (VSLAM) algorithm engineered for high-performance real-time applications. Designed to handle both monocular and stereo camera inputs across varying frame-rates, it aims to bridge the gap between accuracy, robustness, and real-time (RT) processing capabilities.

Key Contributions

1. Multi-threaded Architecture: OV$^{2}$SLAM utilizes a refined four-thread architecture comprising a visual front-end, mapping, state optimization, and loop closure, optimizing both computation and efficiency.
2. Adaptive Localization: This algorithm manages an array of visual localization challenges by incorporating recent advances and enforces real-time constraints, essential for practical applications like autonomous navigation and augmented reality.
3. Online Bag-of-Words (BoW) Integration: Unique to OV$^{2}$SLAM is its use of an online BoW approach (iBoW-LCD) to perform loop closure. Unlike pre-trained BoWs, this method constructs vocabulary incrementally, allowing the system to adapt to diverse environments dynamically.

Numerical and Experimental Insights

The paper rigorously evaluates the system on several benchmark datasets: EuRoC, KITTI, and TartanAir. Results indicate OV$^{2}$SLAM's superior performance in scenarios with real-time demands:

• Versus ORB-SLAM: OV$^{2}$SLAM outperforms ORB-SLAM significantly when evaluated under real-time constraints, showing lower ATE and RPE metrics on EuRoC and KITTI datasets.
• High-frequency Processing: Demonstrates capability in handling frame rates up to 400 Hz, indicating its applicability in high-speed environments.
• TartanAir Benchmark: On this challenging synthetic dataset, OV$^{2}$SLAM achieves robust localization capabilities, particularly outperforming traditional VSLAM methods like ORB-SLAM.

Algorithmic Design

The algorithm addresses several key challenges:

• Keypoint Tracking and Pose Estimation: Implements a guided optical flow method for efficient keypoint tracking and uses nonlinear optimization for pose refinement.
• Temporal and Stereo Matching: Provides robust solutions for both monocular and stereo setups, enhancing the 3D map accuracy through effective triangulation strategies.
• Local Bundle Adjustment (BA): Leverages anchored points with inverse depth representation to streamline optimization without compromising on precision.
• Loop Closure: Employs a loose bundle adjustment strategy to incrementally correct the map without extensive computational overhead.

Implications and Future Directions

The advancements showcased by OV$^{2}$SLAM have substantial implications for real-world applications requiring adaptability and efficiency in dynamic environments. Its robust performance across diverse datasets suggests further exploration could enhance autonomous driving systems, drone navigation, and augmented reality devices.

Future research could focus on:

• Expanding Multi-modal Capabilities: Integrating additional sensor data, such as LIDAR or IMU, to improve robustness in extreme situations lacking visual features.
• Enhanced Loop Closure Techniques: Refining online BoW methods to improve relocalization in continually changing environments.
• Scalability: Ensuring the algorithm maintains high accuracy in even larger and more complex environments, representative of real-world urban landscapes.

In conclusion, OV$^{2}$SLAM sets a high standard for VSLAM technologies, uniquely balancing the trade-offs between performance, robustness, and real-time processing. Its comprehensive design and extensive evaluation offer a promising step forward in the deployment of intelligent visual systems across various demanding applications.