This master's thesis presents the design, implementation, and evaluation of a perception payload for autonomous small-UAS navigation in structured, GNSS-denied environments such as maintenance tunnels and indoor corridors. The system integrates depth imagery, Short-Wave Infrared (SWIR) imagery, and Long-Wave Infrared (LWIR) thermal imaging with an Inertial Measurement Unit (IMU) to simultaneously estimate the platform's 6-DOF pose and generate a thermal anomaly map of the environment.
Small-size UAS (sUAS) offer unique access to hazardous and confined environments — such as underground steam tunnels and industrial maintenance shafts — where GNSS signals are entirely unavailable and human inspection is risky. Achieving reliable autonomous navigation in these unstructured, GPS-denied spaces requires a carefully designed perception payload and a computationally efficient navigation algorithm capable of running in real time on a low-power embedded processor.
The perception payload was integrated on a Blackout quadcopter equipped with:
All sensors were interfaced and time-synchronized via ROS on the Odroid XU4.
A rigorous calibration pipeline was developed covering intrinsic calibration (lens distortion, focal length) for both the Occipital and Lepton cameras, extrinsic calibration (rotation and translation) between the two imaging sensors using a shared asymmetric circle grid target, and temporal calibration between the IMU and the point cloud stream using cross-correlation of orientation signals. Calibration results showed sub-degree rotation error between the cameras.
A RANSAC-based planar surface extraction algorithm was developed to identify floor, ceiling, and wall planes from the incoming 3D point cloud. The algorithm uses environment slicing — partitioning the point cloud into horizontal z-slices — to decompose the scene into 2D cross-sections from which lines, points, and arc features are extracted and combined into 3D planar surfaces. The algorithm runs in real time at 3 Hz on the Odroid XU4 (40–80 ms per frame).
A thermal feature extraction pipeline was implemented to automatically detect hot pipes, steam leaks, and insulation failures from LWIR imagery. The approach uses dynamic histogram-based thresholding (normalizing for ambient temperature variation) and edge extraction to isolate thermally significant objects. Detected anomalies are projected into the 3D map for geo-referenced inspection reporting. Results from Ohio University's steam tunnels successfully identified hot pipes and a steam leak with condensed water on the floor.
The hierarchical Guidance, Navigation and Control (GNC) framework integrates the planar surface features from the 3D imager with the IMU in a Kalman filter for 6-DOF pose estimation. A Complementary Kalman Filter (CKF) drives the dynamics estimator, and an Extended Kalman Filter (EKF) handles feature-based position updates. The planar and point feature observations constrain translation and heading drift during GPS-denied flight.
The platform was flight-tested in Ohio University's underground maintenance steam tunnels and in Stocker Center hallways. Key results include: