Degradation-Aware LiDAR-Thermal-Inertial SLAM

Why Degradation Awareness Matters

LiDAR-inertial and visual-inertial SLAM systems often work well in normal scenes, but disaster environments break their assumptions in different ways. Smoke and dust scatter optical measurements, darkness removes RGB texture, and long corridors or open areas can reduce LiDAR geometric constraints. DaLiTI treats these failures as a sensing-quality problem rather than assuming every modality is equally reliable at every moment.

The paper targets robotic disaster relief settings where GNSS may be unavailable and the robot still needs continuous state estimation. Thermal infrared sensing is useful because heat radiation remains observable in darkness and smoke, but thermal images also have low contrast and high noise. The system therefore combines LiDAR, thermal, and IMU measurements with a mechanism that can decide how much each modality should influence the state update.

Adaptive LiDAR-Thermal-Inertial Fusion

DaLiTI uses an iterated error-state Kalman filter to loosely integrate IMU propagation, LiDAR measurements, and thermal-inertial odometry. The loose structure is deliberate: it keeps the LiDAR and thermal streams separable enough to down-weight a degraded source before its failure propagates through the whole estimator.

The key module is a degradation quantizer. For LiDAR, quality is estimated from the effective plane-feature density. For thermal imagery, quality is estimated from the density of usable thermal edge features. The normalized degradation factors then act as adaptive weights in the IESKF update: LiDAR dominates in geometrically reliable scenes, while thermal observations receive more weight when smoke or other degradation reduces LiDAR reliability.

Thermal Features with LiDAR-Aided Depth

The thermal branch follows the edge-based thermal odometry idea from ETIO: raw thermal images are converted into edge images so that data association depends on thermal-radiation boundaries rather than low-contrast intensity patches. LiDAR is then used to provide depth for thermal edge features when the LiDAR quality is high enough.

This cross-modality association avoids blindly averaging nearby projected LiDAR points around an image edge, where foreground and background points often mix. Instead, the method filters background points and only trusts LiDAR-aided depth when the degradation factor indicates that LiDAR perception is reliable.

Fire and Chemical-Plant Experiments

The system was evaluated on a wheeled robot with an Ouster OS1-64 LiDAR, Xenics Gobi+ 640 thermal camera, Xsens MTi-G-710 IMU, and RGB-D camera. Experiments ran on an Intel i7-11800H laptop without GPU acceleration.

In the indoor fire scenario, the robot traversed a textureless corridor, entered a dark smoke-filled warehouse, and returned along an approximately 110 m trajectory. DaLiTI achieved less than 0.1 m accumulated drift, while most comparison methods failed or accumulated much larger error inside the smoke-filled area.

For the outdoor chemical-plant leakage scenario, the robot moved at night around a smoke-obscured water tank over a 60 m trajectory. The complete DaLiTI system achieved the best trajectory accuracy among the evaluated variants and baselines, showing the benefit of using thermal observations without forcing unreliable measurements into the estimator.

Open-Source Release

The implementation is released as an open-source contribution through HITSZ nROS-Lab. The project connects directly to the published RA-L paper and the public GitHub repository, so the webpage keeps both the research-paper context and the live open-source status visible.