一句话：提出领域特定运行时调度器 SlimSLAM，按输入需求动态调节 SLAM 参数，在保持精度的同时降低计算量。
视觉-惯性 SLAM 需实时且高精度，但计算开销大，受限平台难以兼顾。现有算法多按最坏情况静态配置，导致计算预算利用不当。本文提出 SlimSLAM，一种领域特定的运行时调度器，利用 SLAM 状态信息实时检测并调整过配置的算法参数，以输入需求为导向最小化计算量。作者在 HybVIO 上实现并评估，在 EuRoC、TUM-VI Room 与 Monado VR 数据集上最高获得 5.4× 加速，在同精度下平均优于其他自适应方法 2.3×–1.3×；在 Raspberry Pi 4 上比静态配置及其他 SLAM 算法更快且更准确。
Simultaneous localization and mapping (SLAM) algorithms track an agent's movements through an unknown environment. SLAM must be fast and accurate to avoid adverse effects such as motion sickness in AR/VR headsets and navigation errors in autonomous robots and drones. However, accurate SLAM is computationally expensive and target platforms are often highly constrained. Therefore, to maintain real-time functionality, designers must either pay a large up-front cost to design specialized accelerators or reduce the algorithm's functionality, resulting in poor pose estimation. We find that most SLAM algorithms are statically configured for the worst-case input and mismanage their computation budget. Thus, we present SlimSLAM, a domain-specific runtime scheduler which adapts SLAM algorithmic parameters based on input needs, minimizing computation while maintaining accuracy. SlimSLAM exploits information from a SLAM algorithm's state to detect and adjust over-provisioned parameters in real-time. We demonstrate SlimSLAM on the state-of-the-art real-time visual-inertial SLAM algorithm: HybVIO. We show that SlimSLAM is able to provide speedups up to 5.4× on the EuRoC, TUM-VI Room, and Monado VR datasets and outperforms other adaptive approaches on average by 2.3×-1.3× with iso-accuracy. This enables more accurate SLAM on constrained computation platforms such as the Raspberry Pi 4 (RPi4) where SlimSLAM is faster and more accurate than both HybVIO's static RPi4 configuration as well as other SLAM algorithms.
@inproceedings{BehrooziCFSSM24,
title = {SlimSLAM: An Adaptive Runtime for Visual-Inertial Simultaneous Localization and Mapping},
author = {Armand Behroozi and Yuxiang Chen and Vlad Fruchter and Lavanya Subramanian and Sriseshan Srikanth and Scott Mahlke},
booktitle = {ASPLOS},
year = {2024},
doi = {10.1145/3620666.3651361}
}