Edge Computing Offloading Strategy for ORB-SLAM3 Mapping Algorithm

0
0
0

Edge Computing Offloading Strategy for ORB-SLAM3 Mapping Algorithm

Introduction

Simultaneous Localization and Mapping (SLAM) is a fundamental technology that enables robots and autonomous systems to navigate unknown environments by constructing maps while simultaneously determining their position within these maps. Visual SLAM (VSLAM), which relies on visual sensors such as cameras, has gained significant attention due to its cost-effectiveness compared to other SLAM approaches like LiDAR-based systems. Among VSLAM techniques, ORB-SLAM3 stands out as one of the most advanced frameworks, supporting monocular, stereo, and RGB-D cameras while incorporating visual-inertial SLAM capabilities, improved scene recognition, and multi-map mechanisms.

Despite its advancements, ORB-SLAM3 faces challenges related to high computational resource consumption, particularly during the mapping and optimization phases. This poses a significant burden on mobile devices with limited processing power. To address this issue, edge computing has emerged as a promising solution, allowing computationally intensive tasks to be offloaded to nearby edge servers. However, existing edge-assisted SLAM approaches often assume unlimited network resources or rely on heuristic methods for offloading decisions, leading to performance degradation in resource-constrained scenarios.

This paper introduces an enhanced ORB-SLAM3 algorithm that integrates an edge computing offloading strategy to optimize computational efficiency while maintaining high localization accuracy. The proposed method leverages dynamic programming for keyframe selection, uncertainty quantification models, and Mahalanobis distance minimization to improve both local and global mapping processes. Experimental results demonstrate significant improvements in accuracy and computational efficiency compared to traditional ORB-SLAM3.

Background and Motivation

ORB-SLAM3 builds upon its predecessor, ORB-SLAM2, by introducing several key enhancements, including support for visual-inertial SLAM, a more robust place recognition system, and a multi-map architecture. The system consists of multiple parallel threads: tracking, local mapping, and loop closing. The local mapping thread plays a crucial role in maintaining the local map, optimizing keyframe poses, and reducing accumulated errors. However, this process demands substantial computational resources, making it challenging for mobile devices to execute efficiently.

Edge computing offers a potential solution by distributing computational tasks between mobile devices and edge servers. By offloading resource-intensive operations, such as global map optimization, to edge servers, mobile devices can reduce their computational load while maintaining real-time performance. However, existing edge-assisted SLAM solutions often lack theoretical foundations for optimal offloading decisions, particularly under fluctuating network conditions.

The proposed algorithm addresses these limitations by introducing a systematic approach to keyframe selection and uncertainty minimization. The goal is to achieve high-precision SLAM performance while operating within the constraints of mobile device resources.

Methodology

The proposed algorithm enhances ORB-SLAM3 by incorporating an edge computing offloading strategy that optimizes both local and global mapping processes. The system architecture involves a mobile device equipped with a camera and IMU, which communicates bidirectionally with an edge server. The key components of the methodology include:

  1. Keyframe Subset Selection Using Dynamic Programming

A critical challenge in SLAM is selecting an optimal subset of keyframes for map construction without overwhelming computational resources. The proposed algorithm formulates this problem as a subset function optimization task and employs dynamic programming to find an efficient solution.

The algorithm represents keyframe relationships using a weighted graph, where nodes correspond to keyframes and edges represent relative pose measurements. The selection process aims to minimize uncertainty in pose estimation while adhering to computational constraints. By iteratively evaluating keyframe combinations, the algorithm identifies subsets that provide the most significant reduction in uncertainty.

  1. Uncertainty Quantification and Minimization

To improve mapping accuracy, the algorithm introduces an uncertainty quantification model that evaluates the reliability of keyframe pose estimates. The model assesses the covariance of pose estimations and uses the Mahalanobis distance to optimize the map. By minimizing uncertainty, the algorithm enhances both local and global map consistency.

In the local mapping phase, the mobile device selects keyframes from a candidate pool and constructs a local map. Fixed keyframes, previously optimized on the edge server, are incorporated to anchor the local map and reduce uncertainty. The global mapping phase involves offloading selected keyframes to the edge server, where a comprehensive optimization process refines the global map.

  1. Edge-Assisted Map Optimization

The algorithm partitions computational tasks between the mobile device and the edge server to balance efficiency and accuracy. The mobile device handles real-time tracking and local map optimization, while the edge server manages loop closure detection and global map refinement. This division ensures that the mobile device operates within its computational limits while leveraging the edge server’s processing power for complex optimizations.

Experimental Evaluation

The performance of the proposed algorithm was evaluated using the TUM RGB-D dataset, a standard benchmark for SLAM systems. The experiments compared the enhanced algorithm against traditional ORB-SLAM3 in terms of localization accuracy and computational efficiency.

  1. Localization Accuracy

The evaluation metrics included Absolute Trajectory Error (ATE) and Relative Pose Error (RPE), which measure global and frame-to-frame accuracy, respectively. The results demonstrated that the proposed algorithm achieved higher precision across multiple datasets, with an average improvement of 14.2% in localization accuracy.

  1. Computational Efficiency

CPU usage was measured to assess the algorithm’s resource efficiency. The experiments revealed that the enhanced algorithm reduced average CPU consumption by 20.7% compared to ORB-SLAM3. This reduction was attributed to the dynamic programming-based keyframe selection and the offloading of intensive computations to the edge server.

Discussion

The experimental results validate the effectiveness of the proposed edge computing offloading strategy for ORB-SLAM3. By intelligently selecting keyframes and minimizing uncertainty, the algorithm achieves higher accuracy with fewer computational resources. The dynamic programming approach ensures that keyframe selection is both efficient and optimal, while the uncertainty quantification model enhances map consistency.

One notable advantage of the algorithm is its adaptability to resource-constrained environments. Unlike traditional SLAM systems that require high-end hardware, the proposed method enables reliable SLAM performance on mobile devices by leveraging edge computing. This makes it particularly suitable for applications such as autonomous drones, augmented reality, and mobile robotics.

Conclusion

This paper presented an enhanced ORB-SLAM3 algorithm that integrates an edge computing offloading strategy to improve computational efficiency and localization accuracy. By employing dynamic programming for keyframe selection, uncertainty minimization techniques, and edge-assisted optimization, the algorithm achieves superior performance compared to traditional ORB-SLAM3.

The experimental results demonstrated significant improvements in both accuracy and resource utilization, making the algorithm a viable solution for real-world applications with limited computational capabilities. Future work could explore further optimizations in keyframe selection and uncertainty modeling, as well as extensions to multi-robot collaborative SLAM scenarios.

doi.org/10.19734/j.issn.1001-3695.2024.05.0147

Was this helpful?

0 / 0