Deep feature-enhanced LiDAR-visual-inertial odometry for robust mapping in tunnel environment

Tunnel operation and maintenance underpin infrastructure system functionality. With intelligent automation, robotic tunnel inspection and mapping are gradually replacing manual methods. Odometry is a core technology for robotic inspection, but sparse geometric features and repetitive structures in tunnels cause LiDAR odometry (LO) performance degradation. This paper proposed a LiDAR-vision-inertial odometry fusion framework for tunnel-degraded environments. It adopts coupled rotation-translation constraints to identify 6-DoF constraint degradation. Deep neural network-based visual-inertial odometry (VIO) is proposed to enhance feature tracking robustness in low-texture scenes. A Conditional Extended Kalman Filter (CEKF) fusion architecture selectively integrates VIO information along degradation directions. Experiments on two authoritative tunnel datasets yield Absolute Trajectory Error (ATE) of 3.23 m and 3.74 m, Mean Map Entropy (MME) of −8.58 and − 7.04, with 16.14% Relative Pose Error (RPE) improvement over the BALM, while maintaining real-time performance. This paper provides a comprehensive solution for robot inspection in tunnel-degraded environments. • LiDAR odometry degradation detection method • Deep neural network-based visual-inertial odometry for feature matching • Conditional Extended Kalman Filter-based data fusion architecture • Mean Map Entropy (MME) of −8.58 and −7.04 on respective datasets • Robust mapping performance in tunnel environments