Technical Summary: A Modular Dual-Camera Pipeline for Micro-Inspection Using Aerial Robots

Problem Statement

Existing aerial inspection systems face significant challenges when tasked with inspecting small targets (micro-inspection). Current approaches often require drones to fly dangerously close to targets or follow complex flight paths to capture fine details, increasing collision risks and operational costs. Furthermore, drone flight is susceptible to environmental disturbances and localization inaccuracies. When a drone relies on a narrow field-of-view (FOV) camera for inspection, these inaccuracies can cause the system to lose sight of the target entirely. Additionally, many existing trajectory planning methods depend on prior knowledge of the target's geometry, position, and orientation—information that is frequently unavailable for non-structural targets like trees, vehicles, or people.

Methodology

The paper presents aerial_micro_inspection, a generic, open-source pipeline built on ROS 2 and compatible with the PX4 flight stack. The system utilizes a dual-camera architecture mounted on an aerial robot:

1. Wide-FOV Stereo Navigation Camera: Rigidly attached to the drone body, this camera acquires the target surface on-site, estimates depth via stereo vision, and segments the surface into smaller inspection regions (partitions). It also supports state estimation and navigation.
2. Narrow-FOV Gimbal-Mounted Inspection Camera: Equipped with zoom optics, this camera captures high-detail images of specific partitions without requiring the drone to fly in immediate proximity to the target.

Key Technical Components

• Kinematic Modeling & Calibration: The system employs a kinematic model to map pixels from the navigation camera to the inspection camera. A custom extrinsic calibration tool using an ArUco marker determines the transformation between the rigidly mounted navigation camera and the gimbal base, accounting for pan-tilt degrees of freedom.
• Surface Partitioning: The pipeline uses a parametric sweeping strategy (Algorithm 1) to divide the segmented target surface into partitions that match the projected FOV of the inspection camera. This allows for systematic coverage without prior geometric knowledge.
• Vision-Based Feedback Loop: To counteract drone motion and localization errors, a surface tracking module (using Median Flow) continuously updates the location of partitions in real-time. This feedback adjusts the gimbal's pointing direction, ensuring the inspection camera remains aligned with the intended partition even if the drone drifts.
• Modular Architecture: The pipeline is designed with interchangeable checkpoints for surface segmentation and micro-target detection (e.g., YOLO-based models), allowing adaptation to various use cases.

Key Contributions

The authors identify four primary contributions:

1. First Open-Source Modular Pipeline: To the best of their knowledge, this is the first ROS 2-based pipeline explicitly focused on aerial micro-inspection, offering a modular framework for different applications.
2. Parametric Sweeping Strategy: A method to divide any inspection target into zoom-compliant partitions for systematic coverage.
3. Vision-Based Feedback Mechanism: A control loop that maintains gimbal alignment with intended partitions despite localization errors and platform motion.
4. Extrinsic Calibration Tool: A specific tool for calibrating the relationship between a rigidly attached wide-view camera and a gimbal-mounted inspection camera.

Experimental Results

The pipeline was evaluated in both simulation (Gazebo with PX4 SITL) and real-world experiments across two primary use cases:

1. Tree Inspection: Detecting oak processionary caterpillars and their eggs (1–2 mm) on tree trunks.
2. Greenhouse Inspection: Monitoring sticky traps for whiteflies in a GPS-denied environment.
3. Simulation (Vehicle Inspection): A secondary simulation scenario involving car-body surface inspection to demonstrate transferability.

Performance Findings

• Robustness to Disturbance: In simulation, the surface-tracking feedback loop significantly improved coverage robustness. While open-loop execution suffered from coverage degradation under GPS noise (standard deviation of 1.0 m) and external disturbances, the feedback-enabled system recovered quickly and maintained high coverage.
• Real-World Accuracy:
  • Tree Inspection: Achieved an average angular pointing error of 2.17°, which was sufficient for reliable caterpillar detection.
  • Greenhouse Inspection: Achieved an average error of 3.34°. While this allowed for high-quality imaging of insects, the authors note it was insufficient to distinguish whiteflies from other insects at the current zoom level.
• Limitations Observed: The authors noted that the gimbal's internal zero reference can shift between power cycles, introducing pointing biases that grow with zoom levels (e.g., up to 35% of the image width at 10× zoom).

Significance and Claims

The paper claims that aerial_micro_inspection addresses the critical need for reliable, autonomous micro-inspection that reduces operational risk and personnel exposure. By decoupling the navigation and inspection tasks and integrating a vision-based feedback loop, the system achieves stable close-up imaging from safer distances without requiring prior knowledge of the target geometry.

The authors emphasize that the pipeline is a foundational step toward generalizable aerial inspection. They modestly acknowledge current limitations, such as modeling simplifications regarding gimbal alignment and the need for higher precision at extreme zoom levels. They propose that future work will focus on refining kinematic models, transitioning to online targetless calibration, and expanding the planning layer to handle hierarchical viewpoint generation for larger structures. The code is released as open-source to facilitate adaptation to new domains such as infrastructure monitoring and industrial quality control.