Industry Research

Appearance

Omni-LIVO and Multi-LVI-SAM ​

Related docs: LVI-SAM, FAST-LIVO and FAST-LIVO2, R2LIVE and R3LIVE, LIR-LIVO, multi-sensor robust state estimation, and multi-LiDAR calibration.

Last updated: 2026-05-09

Executive Summary ​

Omni-LIVO and Multi-LVI-SAM are 2025 multi-camera LiDAR-visual-inertial odometry systems. They address a practical weakness in many LIVO systems: a wide-FOV LiDAR observes much more of the scene than a single conventional camera, so visual constraints cover only part of the LiDAR geometry. Multi-camera rigs reduce that mismatch.

Omni-LIVO is a tightly coupled multi-camera LIVO system. It introduces cross-view direct alignment for photometric consistency across non-overlapping views and extends an Error-State Iterated Kalman Filter with multi-view updates and adaptive covariance.

Multi-LVI-SAM extends the LVI-SAM/factor-graph style toward multiple fisheye cameras. It uses a panoramic visual feature model to unify multi-camera observations, adds extrinsic compensation for triangulation consistency, and integrates the panoramic model into a tightly coupled LiDAR-visual-inertial factor graph.

For airside autonomy, multi-camera LIVO is relevant because airport vehicles often need 360-degree perception around tugs, trailers, aircraft, and GSE. It is still a research direction unless calibration, synchronization, dynamic-object rejection, and health monitoring are engineered carefully.

What They Add ​

MethodEstimator styleCore visual ideaPractical value
Omni-LIVOESIKF fusionCross-view direct photometric alignment and adaptive covarianceUses multiple cameras to exploit more LiDAR depth
Multi-LVI-SAMFactor graphPanoramic visual feature model for fisheye camerasUnified multi-camera constraints and loop/global optimization

Inputs and Outputs ​

Inputs:

  • 3D LiDAR.
  • IMU.
  • Multiple cameras, including conventional or fisheye rigs depending on method.
  • Camera intrinsics, distortion models, and camera-LiDAR-IMU extrinsics.
  • Accurate timestamps and exposure metadata.

Outputs:

  • LiDAR-visual-inertial trajectory.
  • Local map or colored point cloud/map products.
  • Multi-view visual feature constraints.
  • Loop or globally optimized trajectory in the graph-based case.

Core Technical Ideas ​

Omni-LIVO focuses on the field-of-view mismatch. A LiDAR may cover a broad 3D region while a single camera covers only a narrow slice. Omni-LIVO uses multiple cameras and cross-view direct alignment so photometric consistency can be maintained even across non-overlapping views. Its ESIKF update incorporates multi-view measurements and adaptive covariance weighting.

Multi-LVI-SAM focuses on unified multi-camera geometry. Instead of treating each fisheye camera as a separate visual front end, it builds a panoramic visual feature model. This representation consolidates multi-view constraints and supports loop closure and global pose optimization. The extrinsic compensation module addresses triangulation inconsistency between individual camera frames and the panoramic frame.

Pipeline ​

  1. Synchronize LiDAR, IMU, and all camera streams.
  2. Propagate state with IMU.
  3. Deskew LiDAR and associate visual information with LiDAR depth where applicable.
  4. Track multi-view visual features or direct photometric residuals.
  5. Fuse geometric LiDAR residuals and visual residuals in an ESIKF or factor graph.
  6. Apply adaptive covariance or compensation for camera-specific geometry.
  7. Update the local map and publish odometry.
  8. In graph-based pipelines, add loop closures and global optimization constraints.

Strengths ​

  • Better spatial coverage than single-camera LIVO.
  • More robust visual constraints around the vehicle.
  • Useful for 360-degree vehicle rigs and complex close-range maneuvers.
  • LiDAR gives metric depth and geometric anchoring.
  • IMU provides high-rate propagation and attitude stability.
  • Fisheye/panoramic modeling is well matched to robots that need near-full-surround awareness.

Failure Modes ​

  • Multi-camera calibration is harder than single-camera calibration.
  • Small extrinsic or timing errors create inconsistent visual constraints.
  • Fisheye distortion and rolling shutter can bias feature triangulation or direct alignment.
  • Cameras remain sensitive to glare, night, rain, lens dirt, motion blur, and exposure shifts.
  • Dynamic objects can produce strong visual features and LiDAR returns.
  • Photometric methods can fail on low texture, reflections, or sudden lighting changes.
  • Compute grows with camera count unless feature selection and scheduling are controlled.

Airside, Indoor, and Outdoor Fit ​

Indoor: Strong in terminals, baggage halls, warehouses, and hangars where multi-camera coverage helps tight navigation and LiDAR provides geometry.

Outdoor: Useful for vehicle rigs with good calibration, but cameras need exposure control, cleaning, weather handling, and robust feature rejection.

Airside: Highly relevant for tug and GSE platforms with multiple blind spots. Useful around aircraft stands, docking, trailer alignment, and terminal-edge driving. The production baseline should still fuse RTK/GNSS, wheel odometry, LiDAR map localization, and health checks. Multi-camera LIVO should degrade gracefully when cameras are blinded by floodlights, rain, spray, or contamination.

Implementation Notes ​

  • Use hardware triggering or audited timestamping for all cameras, LiDAR, and IMU.
  • Calibrate each camera to IMU and LiDAR, then validate cross-camera consistency on real vehicle logs.
  • Track per-camera health: exposure, blur, feature count, residuals, occlusion, and contamination.
  • Avoid letting one bad camera dominate the estimator; use adaptive covariance and camera-level gating.
  • Validate camera overlap with LiDAR across near-field and far-field zones.
  • Use dynamic masks for aircraft, GSE, personnel, reflections, and screens.
  • Benchmark against LVI-SAM, FAST-LIVO2, R3LIVE, and a LiDAR-inertial-only baseline.

Sources ​

Public research notes collected from public sources.

Copyright © 2026 kvynlim