Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Real-Time Thermal-Inertial Odometry on Embedded Hardware for High-Speed GPS-Denied Flight

Austin Stone, Mark Petersen, Cammy Peterson

TL;DR

An uncertainty-aware gated recurrent unit (GRU) network is trained that models the temporal dynamics of static pressure distortion, outperforming polynomial and multi-layer perceptron (MLP) baselines and expanding the operational envelope of thermal-inertial navigation.

Abstract

We present a real-time monocular thermal-inertial odometry system designed for high-velocity, GPS-denied flight on embedded hardware. The system fuses measurements from a FLIR Boson+ 640 longwave infrared camera, a high-rate IMU, a laser range finder, a barometer, and a magnetometer within a fixed-lag factor graph. To sustain reliable feature tracks under motion blur, low contrast, and rapid viewpoint changes, we employ a lightweight thermal-optimized front-end with multi-stage feature filtering. Laser range finder measurements provide per-feature depth priors that stabilize scale during weakly observable motion. High-rate inertial data is first pre-filtered using a Chebyshev Type II infinite impulse response (IIR) filter and then preintegrated, improving robustness to airframe vibrations during aggressive maneuvers. To address barometric altitude errors induced at high airspeeds, we train an uncertainty-aware gated recurrent unit (GRU) network that models the temporal dynamics of static pressure distortion, outperforming polynomial and multi-layer perceptron (MLP) baselines. Integrated on an NVIDIA Jetson Xavier NX, the complete system supports closed-loop quadrotor flight at 30 m/s with drift under 2% over kilometer-scale trajectories. These contributions expand the operational envelope of thermal-inertial navigation, enabling reliable high-speed flight in visually degraded and GPS-denied environments.

Real-Time Thermal-Inertial Odometry on Embedded Hardware for High-Speed GPS-Denied Flight

TL;DR

An uncertainty-aware gated recurrent unit (GRU) network is trained that models the temporal dynamics of static pressure distortion, outperforming polynomial and multi-layer perceptron (MLP) baselines and expanding the operational envelope of thermal-inertial navigation.

Abstract

We present a real-time monocular thermal-inertial odometry system designed for high-velocity, GPS-denied flight on embedded hardware. The system fuses measurements from a FLIR Boson+ 640 longwave infrared camera, a high-rate IMU, a laser range finder, a barometer, and a magnetometer within a fixed-lag factor graph. To sustain reliable feature tracks under motion blur, low contrast, and rapid viewpoint changes, we employ a lightweight thermal-optimized front-end with multi-stage feature filtering. Laser range finder measurements provide per-feature depth priors that stabilize scale during weakly observable motion. High-rate inertial data is first pre-filtered using a Chebyshev Type II infinite impulse response (IIR) filter and then preintegrated, improving robustness to airframe vibrations during aggressive maneuvers. To address barometric altitude errors induced at high airspeeds, we train an uncertainty-aware gated recurrent unit (GRU) network that models the temporal dynamics of static pressure distortion, outperforming polynomial and multi-layer perceptron (MLP) baselines. Integrated on an NVIDIA Jetson Xavier NX, the complete system supports closed-loop quadrotor flight at 30 m/s with drift under 2% over kilometer-scale trajectories. These contributions expand the operational envelope of thermal-inertial navigation, enabling reliable high-speed flight in visually degraded and GPS-denied environments.
Paper Structure (31 sections, 17 equations, 5 figures, 5 tables)

This paper contains 31 sections, 17 equations, 5 figures, 5 tables.

Table of Contents

  1. Introduction
  2. Background
  3. VIO
  4. Thermal-Inertial Odometry
  5. High-Speed Navigation
  6. Barometric Altitude Compensation
  7. Setup
  8. Hardware Configuration
  9. Camera Calibration
  10. Flat-Field Correction Scheduling
  11. Methods
  12. Algorithm Overview
  13. Initialization
  14. IMU Preprocessing
  15. Zero-Velocity Detection and Feature Consistency
  16. ...and 16 more sections

Figures (5)

  • Figure 1: Comparison of degraded and good IR images. Degraded image shows characteristic cross-hatch pattern.
  • Figure 2: Overview of our monocular thermal-inertial odometry pipeline.
  • Figure 3: Example trajectories from flight experiments. The ground truth reference was obtained using onboard GPS measurements.
  • Figure 4: Barometric error vs. time for filtered, polynomial, MLP, and GRU corrections.
  • Figure 5: Mean and STD of test error for polynomial models trained with random sub samples of training data