Adaptive Degeneracy-Aware LiDAR-Inertial-Visual Odometry with Balanced Accuracy and Efficiency

Abstract

LiDAR-inertial-visual odometry (LIVO) has emerged as a promising solution for robust localization in challenging environments. However, achieving an optimal balance between computational efficiency and localization accuracy remains difficult in mixed scenes that contain both well-constrained and geometrically degenerate LiDAR segments. Fixed visual update policies often introduce unnecessary computation, while simply reducing visual participation may degrade accuracy under degeneracy. To address this limitation, we propose a degeneracy-aware LiDAR-inertial-visual odometry framework. The key contribution is an adaptive visual update strategy that dynamically adjusts the visual update frequency based on LiDAR geometric quality, reducing computational overhead in well-constrained segments while enhancing visual participation under severe degeneration. To compensate for potential accuracy loss, we further introduce a residual-weighted voxel plane estimation method for more robust LiDAR geometric modeling, along with a degradation-aware visual point augmentation mechanism that strengthens visual constraints in weakly constrained segments. Experimental results on the M3DGR, Hilti22, and Hilti23 datasets demonstrate that the proposed method achieves a superior accuracy-efficiency trade-off compared to the baseline. On the M3DGR dataset, it reduces the average runtime by 22.3% while maintaining comparable localization accuracy. Results on Hilti22 and Hilti23 further validate its competitive localization performance with improved efficiency in challenging real-world environments.