LDSO: Direct Sparse Odometry with Loop Closure (1808.01111v1)

Abstract: In this paper we present an extension of Direct Sparse Odometry (DSO) to a monocular visual SLAM system with loop closure detection and pose-graph optimization (LDSO). As a direct technique, DSO can utilize any image pixel with sufficient intensity gradient, which makes it robust even in featureless areas. LDSO retains this robustness, while at the same time ensuring repeatability of some of these points by favoring corner features in the tracking frontend. This repeatability allows to reliably detect loop closure candidates with a conventional feature-based bag-of-words (BoW) approach. Loop closure candidates are verified geometrically and Sim(3) relative pose constraints are estimated by jointly minimizing 2D and 3D geometric error terms. These constraints are fused with a co-visibility graph of relative poses extracted from DSO's sliding window optimization. Our evaluation on publicly available datasets demonstrates that the modified point selection strategy retains the tracking accuracy and robustness, and the integrated pose-graph optimization significantly reduces the accumulated rotation-, translation- and scale-drift, resulting in an overall performance comparable to state-of-the-art feature-based systems, even without global bundle adjustment.

• The paper introduces LDSO, which integrates loop closure into direct sparse odometry to mitigate drift and improve long-term SLAM accuracy.
• It leverages a BoW feature framework and Sim(3) pose constraints to enhance repeatability in feature detection, even in feature-sparse environments.
• Rigorous evaluation on public datasets shows reduced Absolute Trajectory Errors, making LDSO promising for autonomous navigation and AR applications.

Direct Sparse Odometry with Loop Closure: An Examination of LDSO

This paper presents an advancement in monocular visual Simultaneous Localization and Mapping (SLAM) systems by merging the principles of Direct Sparse Odometry (DSO) with loop closure detection and pose-graph optimization, termed as LDSO. The work builds on the foundation laid by the DSO approach which operates by utilizing image pixels with sufficient intensity gradients, thus maintaining high performance in featureless environments.

The core contribution of LDSO is the integration of a loop closure mechanism to the DSO framework. This is achieved by leveraging a point selection strategy that emphasizes corner features, thereby improving the repeatability necessary for effective loop closure detection. These repeatable points are utilized to verify loop closure candidates using a traditional feature-based bag-of-words (BoW) framework. The differentiation in approach lies in the adoption of Sim(3) relative pose constraints established through simultaneous minimization of 2D and 3D geometric errors, subsequently integrated with a co-visibility graph derived from DSO’s sliding window optimization.

Numerical Results and Performance Evaluation:

The performance of LDSO was rigorously evaluated across several public datasets, including the TUM-Mono, EuRoC MAV, and KITTI odometry datasets. Particularly notable is the system’s ability to maintain tracking accuracy comparable to feature-based systems, despite operating without global bundle adjustment. The loop closure functionality theoretically addresses the typical odometry drift experienced in long-term operations, judiciously managing rotation, translation, and scale. Among highlighted results, the integration of loop closure into LDSO notably improved trajectory accuracy over the original DSO, as evidenced by reduced Absolute Trajectory Errors on benchmarking datasets.

Theoretical Implications:

The advancements addressed by LDSO extend the theoretical capabilities of SLAM systems by refining the balance between direct and feature-based methods. The work demonstrates that it is indeed feasible to meld dense photometric tracking capabilities with sparse feature recognition in a real-time SLAM solution. Furthermore, the system mitigates the computational burdens commonly associated with global bundle adjustment by optimizing the use of available co-visibility graphs for pose correction.

Practical Implications:

LDSO's framework shows substantial promise in real-world applications, including autonomous navigation, augmented reality, and robotic systems requiring prolonged localization accuracy. The robustness against feature-sparse environments ensures its adaptability across diverse deployment conditions. The system's design advocates for decreased reliance on extensive sensor setups such as stereo rigs or inertial measurement units, a constraint that enforces cost-effectiveness in consumer and industrial applications.

Future Developments:

Future investigations might aim to incorporate photometric bundle adjustment strategies to refine map accuracy further. Moreover, there is potential for advancing map maintenance strategies, critical for long-term stability and the computational efficiency of SLAM systems. Techniques such as the strategic removal of redundant keyframes and 3D points could further enhance LDSO's operation without adversely affecting mapping precision. Moreover, the fusion of depth estimates with neighboring keyframes after loop closure could contribute significantly to the overall geometric accuracy of SLAM solutions.

In conclusion, the presented LDSO method offers a sophisticated integration of loop closure into the direct odometry approach, successfully addressing challenges of accumulated drift and scalability in visual SLAM systems. The methodological innovations and numerical validations demonstrated in this work provide a credible contribution to the ongoing development of robust SLAM frameworks.