Towards practical LiDAR-inertial-visual odometry on resource-constrained platform : robustness evaluation and lightweight design

Abstract

Autonomous navigation systems are increasingly deployed in real-world applications, yet their effectiveness on low-power platforms remains limited by the heavy computational demands of modern LiDAR-Inertial-Visual Odometry (LIVO) systems. This thesis investigates the deployment challenges of state-of-the-art LIVO pipelines under resource-constrained settings and proposes a lightweight solution to bridge the gap between laboratory performance and field deployment. First, an extensive empirical evaluation demonstrates that modern systems like FAST-LIVO2 exhibit strong robustness even under extreme perception degradation, such as in texture-less, dark, and smoke-filled environments. However, the evaluation also reveals that existing frameworks consume excessive computation and memory, restricting their applicability on embedded hardware. To address these limitations, this work introduces a degeneration-aware lightweight LIVO system with two key contributions: (1) a LiDAR-degeneration-guided adaptive visual frame selector, which activates image updates only when necessary; and (2) a dual-layer hybrid mapping structure that reduces memory consumption while maintaining tracking accuracy. Extensive experiments on public datasets and custom collected challenging scenarios demonstrate that the proposed system reduces runtime by over 30% and memory usage by nearly 50%, all while maintaining competitive localization accuracy. The method is successfully deployed on a $100 ARM platform in real time. These results highlight the potential of the proposed approach to enable scalable, efficient, and robust autonomy in constrained environments.