FusionLCD: Visual-LiDAR Loop Closure Detection With Global Visual Prefiltering for Multi-Sensor SLAM

Abstract

FAST-LIVO2 tightly couples LiDAR geometry with direct visual updates in a unified voxel map for accurate real-time mapping and tracking, yet drift can still accumulate over long trajectories and lead to global inconsistency. We therefore propose FusionLCD, a lightweight online loop closure detection module for multi-sensor SLAM that integrates vocabulary-free global visual prefiltering with efficient LiDAR geometric verification. Specifically, we encode each camera keyframe into a compact global descriptor by aggregating SuperPoint features using generalized mean and max pooling (GeM+MAC), spatial pyramid pooling (SPP), and R-MAC-based regional pooling, enabling real-time top-$K$ retrieval of loop candidates. The retrieved candidates are then validated by LiDAR geometric gating. A simplified Scan Context and BEV phase correlation estimate coarse yaw and translation priors to initialize ICP, which runs only on pairs that pass all gates. FusionLCD operates on synchronized visual-LiDAR data, which simplifies integration with multi-sensor SLAM systems. We integrate FusionLCD into FAST-LIVO2, and experiments show that the resulting system detects loop closures reliably, reduces drift, improves global consistency, and maintains real-time performance.