[论文解读] FastLoop: Parallel Loop Closing with GPU-Acceleration in Visual SLAM

Soudabeh Mohammadhashemi, Shishir Gopinath+4 more|arXiv (Cornell University)|Mar 17, 2026

Robotics and Sensor-Based Localization被引用 0

一句话总结

FastLoop 通过应用任务级和数据级并行性来加速 ORB-SLAM3 的回环检测模块，在 EuRoC 和 TUM-VI 上实现显著的加速，同时保持轨迹精度。

ABSTRACT

Visual SLAM systems combine visual tracking with global loop closure to maintain a consistent map and accurate localization. Loop closure is a computationally expensive process as we need to search across the whole map for matches. This paper presents FastLoop, a GPU-accelerated loop closing module to alleviate this computational complexity. We identify key performance bottlenecks in the loop closing pipeline of visual SLAM and address them through parallel optimizations on the GPU. Specifically, we use task-level and data-level parallelism and integrate a GPU-accelerated pose graph optimization. Our implementation is built on top of ORB-SLAM3 and leverages CUDA for GPU programming. Experimental results show that FastLoop achieves an average speedup of 1.4x and 1.3x on the EuRoC dataset and 3.0x and 2.4x on the TUM-VI dataset for the loop closing module on desktop and embedded platforms, respectively, while maintaining the accuracy of the original system.

研究动机与目标

• Motivate and address the high computational cost of loop closure in visual SLAM.
• Redesign the loop closing architecture to exploit parallelism and reduce CPU-GPU transfers.
• Integrate GPU-accelerated pose graph optimization to speed up global consistency corrections.
• Evaluate performance and accuracy gains on desktop and embedded platforms using standard SLAM benchmarks.

提出的方法

• Identify task-level and data-level parallelism in the loop closing pipeline.
• Move Projection Search tasks to run concurrently on the GPU to exploit independence.
• Parallelize data-heavy components (Single Projection Search, Triple Projection Search, Loop Fusion) on GPU.
• Keep keyframes resident on GPU memory to minimize data transfers and reuse memory layouts for multiple kernels.
• Replace CPU-based pose graph optimization with a GPU-accelerated Graphite library and use automatic differentiation for Jacobians.
• Use CUDA with lightweight data wrappers to minimize transfer overhead and employ pinned memory for faster host-device transfers.

实验结果

研究问题

• RQ1 How much speedup can be achieved for the loop closing module when accelerated on the GPU across different datasets and hardware?
• RQ2 Does GPU acceleration of loop closing preserve or improve localization accuracy compared to the baseline ORB-SLAM3?
• RQ3 Which components of the loop closing pipeline benefit most from parallelization, and how does performance scale with map size (poses/edges)?
• RQ4 What are the practical data-transfer and memory-management considerations when offloading loop closure to the GPU?

主要发现

• FastLoop achieves average speedups of 1.4x (EuRoC) and 3.0x–3.7x (TUM-VI) for the loop closing module on desktop and embedded platforms, respectively.
• Loop Fusion and Graph Optimization provide the largest gains, with substantial improvements as graph size increases.
• Graph optimization on the GPU yields up to 4.0x speedup on TUM-VI and shows speedups growing with more poses and edges.
• The trajectory accuracy (ATE RMSE) remains comparable to ORB-SLAM3 across evaluated sequences.
• GPU-based keyframe storage and minimized CPU-GPU data transfers reduce transfer overhead and enable efficient GPU utilization.
• Some sequences with small graphs show limited or no speedup due to transfer overhead dominating computation.