In my research on aerial inspection technologies, I identified a critical gap in the operational capabilities of fixed-wing UAVs. While these platforms offer superior endurance and speed compared to multi-rotor systems, their reliance on specific runways and the complexity of autonomous landings in diverse environments have historically limited their application for large-scale automatic inspections. To address these challenges, I designed and implemented a comprehensive system architecture that integrates advanced control algorithms, robust hardware, and intelligent image processing to enable seamless remote takeoff, autonomous landing, and efficient target detection over vast areas.
My primary motivation was to overcome the key limitations outlined in the literature: the complex and costly nature of vertical takeoff and landing (VTOL) fixed-wing systems, the frequent battery swaps required by multi-rotor drones for long-duration missions, and the lack of unified solutions that address both the flight control and the data processing challenges for wide-area surveillance. My solution focuses on a pure fixed-wing platform combined with a sophisticated, embedded processing system to achieve a high level of autonomy and efficiency.
System Architecture and Hardware Design
The cornerstone of my design is a robust hardware platform built around the Samsung S5P6818 processor, an octa-core Cortex-A53 architecture system. This chip was chosen for its excellent balance of low power consumption and high computational performance, which is critical for an embedded system. It is coupled with a Mali 400 graphics engine to handle intensive image processing tasks. The overall hardware system is bifurcated into two primary components: the airborne system and the ground control station.
The airborne system is an integrated unit containing the flight controller, sensors (GPS, IMU), a high-resolution imaging payload, and the main processing board. For the imaging tasks, I selected the DJI Zenmuse X5S camera, which provides 20.4 MP still images and 4K video, and the FLIR T640 thermal imager for heat-based inspections. The ground control station is powered by an Intel Core i9-10900K processor and 32 GB of RAM to support real-time data reception, high-level mission planning, and system monitoring.
The following table summarizes the key hardware components integrated into my system.
| Component Category | Specific Component | Key Specification / Function |
|---|---|---|
| Main Processor | Samsung S5P6818 | Octa-core Cortex-A53, Low-power, Mali 400 GPU |
| Flight Control Unit | Custom Autopilot | Integration of sensors, control loops, actuator commands |
| High-Resolution Camera | DJI Zenmuse X5S | 20.4 MP, 4K Video, 3-axis gimbal stabilization |
| Thermal Imaging Sensor | FLIR T640 | Infrared thermography for temperature anomaly detection |
| Lidar Sensor | Ouster OS1-16 | 3D point cloud generation for terrain mapping and obstacle avoidance |
| Communication Module | SX1262 LoRa Chip | Long-range, low-power data link for control and telemetry |
| Ground Station CPU | Intel Core i9-10900K | High-performance desktop processor for data processing and mission control |
| Ground Station Memory | Corsair Vengeance LPX 32 GB DDR4 | High-speed RAM for real-time data analysis and visualization |
Control Strategy for Remote Takeoff and Landing
A central challenge I tackled was the autonomous control of the fixed-wing UAV during remote takeoff and landing, especially in the presence of wind disturbances. I developed a closed-loop control architecture that separates the longitudinal (pitch and thrust) and lateral (yaw and roll) control problems. For the lateral control, especially during the critical takeoff and landing phases, I defined a specific heading control law.
For the takeoff phase, the heading control law is designed to keep the aircraft on the runway centerline:
$$ \delta_r = K_{\phi}(\phi_r – \phi) + K_i \Delta S $$
where \( \delta_r \) is the rudder deflection command, \( \phi_r \) is the desired bank angle, \( \phi \) is the current bank angle, \( \Delta S \) is the lateral deviation from the runway centerline, and \( K_{\phi} \) and \( K_i \) are proportional and integral gains, respectively.
For the landing phase, the control law becomes more complex to ensure a smooth flare and touchdown:
$$ \phi_r = K_{\phi}^{\phi}(\phi_r – \phi) $$
$$ \delta_r = -K_r r + K_{\phi}^{*} \phi $$
where \( K_{\phi}^{\phi} \) is the transmission coefficient from yaw to roll, \( r \) is the yaw rate, \( K_r \) is the damping gain for the yaw rate, and \( K_{\phi}^{*} \) is the yaw compensation coefficient for the roll angle.
To further enhance the robustness of the system against wind gusts, I implemented an Active Disturbance Rejection Controller (ADRC). The ADRC treats external disturbances and internal model uncertainties as a total disturbance, which is estimated and compensated for in real-time. The core of the ADRC is an Extended State Observer (ESO):
$$ e = a_1 $$
$$ \dot{a}_1 = a_2 – \alpha_1 f(e, z_1, l) $$
$$ \dot{a}_2 = a_3 – \alpha_2 f(e, z_2, l) $$
$$ \dot{a}_3 = a_3 – \alpha_3 f(e, z_3, l) $$
where \( f(e, z, l) \) represents the internal disturbance, \( \dot{a}_1, \dot{a}_2, \dot{a}_3 \) are the outputs of the ESO, and \( \alpha_1, \alpha_2, \alpha_3 \) are the observer gains. The final control output \( u \) generated by the ADRC is:
$$ u = \frac{(u_0 – a_3)}{b} $$
where \( u_0 \) is the output from the nonlinear error feedback and \( b \) is a known compensation factor. This approach significantly improved the aircraft’s stability during the high-risk phases of takeoff and landing.
Image Processing and Target Detection
For the automatic inspection task, my system relies on advanced computer vision to detect small targets from a high altitude. The core of my detection pipeline is a small target detection algorithm based on target characteristics. I first propose candidate regions using a Feature-based Probabilistic Target Proposal (FPTP) method. To confirm a target, I extract features from these candidates using a Fourier transform-based Histogram of Oriented Gradients (FHOG), which provides rotation invariance and is robust to lighting changes.
The FHOG feature extraction process begins with the gradient calculation. For a given pixel with gradient magnitude \( l \) and orientation \( \nabla f(d) \), the distribution function can be represented by an impulse function:
$$ h(\theta) = \|l\| \delta(\theta – \nabla f(d)) $$
The final FHOG features are derived by applying spatial and normalization convolutions. I define \( P_1 \) as the spatial aggregation kernel and \( P_2 \) as the local normalization kernel. The Fourier coefficient \( \hat{f} \) is used to construct a 2D Fourier basis. The final implemented feature \( F_m \) is given by:
$$ F_m = \frac{\hat{f}_{P_1}}{\|L\|^2_{P_2}} $$
The algorithm’s robustness is further validated by its ability to handle rotated images. If \( W_{k,m} \) is the original FHOG feature and \( W’_{k,m} \) is the feature after a rotation by an angle \( \theta \), the relationship is:
$$ W’_{k,m} = e^{-i\theta(m-k)} W_{k,m} $$
The entire workflow begins with image acquisition by the camera, followed by FPTP-based candidate proposal. Each candidate is then processed to extract FHOG features, which are fed into a classifier to confirm the target. Once confirmed, the system can track the target, and the results, including the target’s location and classification, are transmitted to the ground station for operator review and archival.
System Implementation and Experimental Validation
To validate my design, I established a dedicated experimental site in a remote suburban area. The site was equipped with multiple hangars (nests) to support the ‘remote takeoff and landing’ concept, allowing the fixed-wing UAV to operate over a wide network. The experimental platform was used to conduct X-Plane semi-physical simulation tests and real-world flight tests. The prototype UAV specifications are detailed in the table below:
| Parameter | Value |
|---|---|
| Takeoff Weight | 6000 g |
| Length | 1500 mm |
| Wingspan | 2200 mm |
| Aspect Ratio | 5.8 |
| Wing Area | 0.82 m² |
| Propeller Diameter | 480 mm |
The flight test data from the takeoff and landing phase demonstrated excellent stability. My control system ensured a smooth rotation and climb-out during takeoff, and a controlled flare and touchdown during landing. The performance of the inspection system was quantitatively evaluated by comparing my results with two other methods from recent literature. The comparative data, focused on positional accuracy, is shown below:
| Measurement Point | Research Method | Average Error (m) | ||
|---|---|---|---|---|
| X | Y | Z | ||
| 1 | Literature [1] | 0.35 | 0.34 | 0.41 |
| Literature [2] | 0.29 | 0.32 | 0.40 | |
| My Method | 0.15 | 0.11 | 0.12 | |
| 2 | Literature [1] | 0.29 | 0.31 | 0.61 |
| Literature [2] | 0.23 | 0.24 | 0.24 | |
| My Method | 0.11 | 0.24 | 0.24 | |
| 3 | Literature [1] | 0.32 | 0.72 | 0.74 |
| Literature [2] | 0.26 | 0.25 | 0.65 | |
| My Method | 0.11 | 0.14 | 0.11 | |
The results clearly demonstrate that my system achieves superior positional accuracy, with significantly lower average errors in all three axes. This precision is a direct result of the integrated control and image processing strategies. The efficiency of the inspection process, measured as the success rate of target detection within a defined center-location error threshold, was also evaluated.
My method achieved a detection efficiency of 94.2%. In comparison, the methods from Literature [1] and Literature [2] yielded efficiencies of 84.2% and 73.7%, respectively. This represents a substantial improvement of approximately 11.9% in overall task efficiency and coverage range.
To ensure the system is deployable in real-world, often harsh, environments, I conducted rigorous performance verification tests. The results are summarized in the table below, confirming the robustness and reliability of my design.
| Performance Metric | Measured Value / Specification |
|---|---|
| EMI (Electromagnetic Interference) Hardness | 10 V/m |
| ESD (Electrostatic Discharge) Hardness | 4 kV |
| Surge Protection Level | 10 kA |
| False Detection Rate | < 5% |
| Average Recovery Time | 5 s |
| Compatibility Test Pass Rate | > 95% |
| Average Concurrent Processing Time | < 0.14 s |
| Data Transfer Rate | > 0.74 Gbps |
Conclusion and Reflections
In this research, I have successfully designed and implemented a large-scale automatic inspection system for fixed-wing UAVs. My key contributions include a robust embedded hardware architecture based on the Cortex-A53 processor, an ADRC-enhanced control system for reliable remote takeoff and landing in windy conditions, and an intelligent image processing pipeline utilizing FHOG features for accurate small target detection. The experimental results, including an 11.9% improvement in efficiency and high system reliability metrics, validate the effectiveness of this approach. This work provides a practical, high-performance solution for complex, wide-area inspection tasks, effectively addressing the core limitations of current fixed-wing UAV technology and paving the way for its expanded use in various industrial and public safety domains.