SLAM with LiDAR and Inertial Fusion

SLAM=Simultaneous Localization and Mapping enables robots to construct env maps while self-localizing. LiDAR=Light Detection and Ranging provides high-res 3D point clouds (PCs) w/ mm-cm accuracy. IMU=Inertial Measurement Unit delivers high-freq (100–1000 Hz) acc & ang vel via accel & gyro. Fusion of LiDAR+IMU enhances robustness in degenerate envs (e.g., low-texture, dynamic obs). Core challenge: spatio-temporal misalignment b/w sensors; requires ext calib (extrinsic params: R,t b/w LiDAR & IMU; temporal offset Δt). LiDAR measures geometry; IMU captures dynamics but drifts due to integration. Fusion occurs at data, feature, or pose graph level. MSCKF=Multi-State Constraint Kalman Filter uses IMU states to constrain visual frames. LiDAR-IMU SLAM typically uses EKF=Extended Kalman Filter, UKF=Unscented Kalman Filter, or optimization-based BA=Bundle Adjustment. Key frameworks: LIO=Lidar-Inertial Odometry (tightly coupled), LIMO=Lidar-IMU odometry, and LIVOX-SLAM. LIO-SAM fuses LiDAR, IMU, GPS via factor graph opt. Preintegration theory (Forster et al.) enables IMU error modeling b/w LiDAR scans. IMU preint yields delta-pos, vel, ori; used as prior in opt. LiDAR scan registration: ICP=Iterative Closest Point, NDT=Normal Distributions Transform, or LOAM=LiDAR Odometry and Mapping. LOAM extracts edge & planar feats; LIO-NDT uses NDT cells w/ IMU init. LiDAR feats projected into IMU frame using motion est from IMU. Temporal sync critical: hw triggers or sw time alignment via interpolation. Degenerate motion (e.g., pure rotation, linear traj) causes observability loss; IMU aids by providing gravity ref & dynamic motion cues. Gravity vector fixes roll/pitch; mag not needed if IMU w/ good bias est. IMU must be calibrated: bias, scale, misalignment; online est preferred. LiDAR motion distortion: scans taken over dt (e.g., 100ms); IMU enables motion comp by interpolating transform b/w timestamps. Key math: SE(3) Lie group for pose rep; adjoint for error trans; IMU error state in SO(3)×R⁶. Error state vec: δx = [δθ, δv, δp, δbₐ, δbω]ᵀ. IMU meas model: acc_in_body → acc_in_world via rot; integr w/ bias subtr. Discrete preint: Δv = Σ(Rₖ⋅(aₖ−bₐ))⋅Δt, similarly Δp, ΔR. Jacobians of preint w.r.t. bias used in opt. LiDAR residual: min dist b/w curr scan feats & ref map (planes/edges). Factor graph: nodes=poses, factors=constraints (LiDAR reg, IMU prior, loop closures). Optimization: Ceres, GTSAM, or g2o. Loop closure: scan-to-map matching via ICP or desc (e.g., LOAM, ScanContext). Desc-based: FPFH, SHOT, or learned (e.g., DeepLIO). Degeneracy handling: observability-aware filtering; fix unobservable dir via IMU. Failure modes: high dyn obs (e.g., traffic), sensor occlusion, vibration-induced IMU noise. Vibration: high-freq IMU noise → bias instab; use adaptive noise cov or filter. Dynamic obj: mask moving objs via PC diff or motion seg. Outlier rejection: stat (e.g., RANSAC), Mahal dist, or NN-based. Real-time perf: LiDAR proc O(n); IMU preint O(1); graph opt O(m²) w/ m poses; marginalization used (e.g., slide window). LIO-SAM uses sliding window w/ IMU factors retained. Map rep: voxel grid, surfel, or feature-based. Voxel: efficient for NDT; surfel: oriented point; feature: edge/planar lines. Sensor config: LiDAR-IMU extrinsic must be stable; temp changes affect calib. Auto-calib: online ext para est in opt (e.g., as factor). Benchmarks: KITTI, EuRoC MAV, Newer College, UrbanNav. Metrics: ATE=RMS pos err, RPE=rel pose err, CE=closure err. SOTA: FAST-LIO2 (tightly coupled, EKF, direct), R2LIVE (LiDAR-IMU-Vision), VILENS (variational IMU preint). FAST-LIO2 uses iterated Kalman filter; direct PC mapping; no ext feat ext. R2LIVE fuses camera for texture & robustness. Learning-based: DeepLO (NN for motion est), but less interp. Future: neuromorphic IMU, event-LiDAR fusion, 4D spatiotemporal slam (inc time). Edge AI: deploy on Jetson, FPGA; quantize NNs. Pitfalls: poor extrinsic calib → drift; unmodled IMU bias → div; lack of loop closure → no global cor; over-marginalization → info loss. Best practices: use high-qual IMU (low noise, bias drift), precise time sync, online calib, adaptive thresholds, and multi-sensor fusion (e.g., +GPS). Domain apps: autonomous vehicles, UAVs, indoor drones, planetary rovers.