Proactive Feature Quality Assessment for LiDAR-Inertial Odometry Using Physical Signal Analysis
作者
Niusha Khosravi,Rodrigo Ventura,Meysam Basiri
出处
期刊:IEEE International Conference on Autonomous Robot Systems and Competitions [Institute of Electrical and Electronics Engineers] 日期:2026-04-22卷期号:: 138-144
标识
DOI:10.1109/icarsc70216.2026.11523258
摘要
Most traditional LiDAR-Inertial Odometry (LIO) frameworks rely on uniform measurement noise models, which can limit robustness in geometrically degenerate or densely cluttered environments characterized by varying signal reliability. Although adaptive methods exist to mitigate this issue, they are predominantly reactive, adjusting covariance only after analyzing optimization residuals. Consequently, unreliable measurements may already have affected the state estimate before they are identified and downweighted. In this work, we propose a proactive feature quality assessment framework that estimates measurement reliability a priori using three physical signal cues: photometric stability, quantified through local intensity variance; geometric planarity, derived from eigenvalue-based structure analysis; and range-dependent uncertainty. These cues are fused into a per-point trust score that modulates the measurement covariance within the FAST-LIO2 Iterated Extended Kalman Filter (IEKF). The framework uses physically interpretable parameters, although parameter selection remains sensor-dependent. We evaluate the method on the MARS-LVIG benchmark and on a custom aerial golf-course dataset. On the MARS-LVIG featureless sequence, the proposed method reduces trajectory error by 63.1% relative to FAST-LIO2, with only 4% runtime overhead. On the golf-course dataset, the method yields qualitatively improved map consistency in vegetation-dominated regions, although this second evaluation remains preliminary because RTK-GPS ground truth was not available. Overall, the results indicate that proactive, physics-informed measurement weighting is a promising direction for improving robustness in challenging aerial LiDAR environments.