coVoxSLAM: GPU Accelerated Globally Consistent Dense SLAM

Abstract: A dense SLAM system is essential for mobile robots, as it provides localization and allows navigation, path planning, obstacle avoidance, and decision-making in unstructured environments. Due to increasing computational demands the use of GPUs in dense SLAM is expanding. In this work, we present coVoxSLAM, a novel GPU-accelerated volumetric SLAM system that takes full advantage of the parallel processing power of the GPU to build globally consistent maps even in large-scale environments. It was deployed on different platforms (discrete and embedded GPU) and compared with the state of the art. The results obtained using public datasets show that coVoxSLAM delivers a significant performance improvement considering execution times while maintaining accurate localization. The presented system is available as open-source on GitHub https://github.com/lrse-uba/coVoxSLAM.

• The paper introduces a novel GPU-accelerated dense SLAM approach that achieves real-time, globally consistent mapping.
• It employs optimized TSDF integration and GPU-based pose graph optimization, realizing speed-ups up to 140x over existing methods.
• The system maintains high accuracy with low RMSE on large-scale LiDAR data, proving its practical efficiency for mobile robotics.

An Overview of the coVoxSLAM System: GPU-Accelerated Dense SLAM

The paper "coVoxSLAM: GPU Accelerated Globally Consistent Dense SLAM" by Emiliano Höss and Pablo De Crist introduces a novel approach to Simultaneous Localization and Mapping (SLAM) with a focus on enhancing computational efficiency through GPU acceleration. The coVoxSLAM system builds on existing SLAM methodologies by integrating GPU processing capabilities to achieve real-time map building in large-scale and complex environments, maintaining global consistency without sacrificing accuracy.

Introduction and Motivation

The authors articulate the critical role of dense SLAM systems in enabling mobile robots to navigate unstructured environments autonomously. SLAM systems estimate the trajectory of a robot while incrementally constructing a map of the observed area. However, traditional systems face challenges related to computational demands, especially when maintaining globally consistent and accurate maps over extended areas, requiring substantial resources for loop closing and pose estimation. The transition towards utilizing GPUs in dense SLAM responds to these increasing computational demands by leveraging their inherent parallel processing capabilities.

System Architecture

coVoxSLAM’s architecture, as detailed in the paper, consists of two primary components: the frontend and the backend, both fully optimized to harness GPU capabilities. The frontend focuses on integrating sensor data into a Truncated Signed Distance Field (TSDF) using improved raycasting techniques rather than traditional projection mapping. The transition to raycasting allows for a consistent computational workload distribution across GPU threads, thereby significantly enhancing efficiency. The backend involves processing pose graph optimization entirely on the GPU to refine submap alignment, leveraging iterative methods like Conjugate Gradient and Levenberg-Marquardt for cost function minimization.

Numerical Evaluation and Results

The authors conduct a thorough evaluation of coVoxSLAM compared to existing systems, such as Voxgraph and nvBlox. The numerical results highlight that coVoxSLAM demonstrates a marked improvement in execution times. For TSDF integration over large-scale LiDAR datasets, the system achieves speed-ups ranging from 30x to 140x compared to Voxblox. Notably, it is reported to outperform nvBlox with a performance increment of up to 2x in similar scenarios. Additionally, coVoxSLAM effectively manages to maintain accuracy levels akin to previous state-of-the-art methods, as evidenced by root mean square error (RMSE) comparisons against ground truth trajectories.

Implications and Future Directions

The introduction of coVoxSLAM contributes significantly to the SLAM domain by proving that complex volumetric mapping can meet the demands of real-time processing in larger and more intricate environments without the burdensome computational cost. The system's ability to function efficiently on both discrete and embedded GPU architectures highlights its adaptability and practical relevance, offering substantial utility for on-board applications in mobile robotics and beyond.

Looking ahead, the integration of deep learning techniques with coVoxSLAM presents an intriguing opportunity for further enhancing the quality and robustness of dense SLAM, especially in feature extraction and semantic mapping. Moreover, as GPU architectures continue to evolve, SLAM systems like coVoxSLAM could benefit from increased computational power and improved parallel processing models, paving the way for more advanced autonomous navigation solutions across diverse sectors.

In conclusion, the coVoxSLAM system delineated in this paper serves as a testament to the potential of GPU acceleration in addressing complex challenges in dense SLAM. The work sets a new benchmark in the domain, inviting further research to expand on its capabilities and explore its full potential in broader applications.