Research on Precise Landing of Rotor UAV Based on Multi-Sensor Fusion

Jintang LiInternational Journal of Computer Science and Information Technology 2024 · DOI

Abstract

With the rapid development of drone technology, the autonomous landing of drones has become a current research hotspot. This article introduces an autonomous drone landing system based on multi-sensor fusion. The system combines GPS, IMU, visual sensors, LiDAR, and other sensor data through Kalman filtering to achieve data fusion and improve the positioning accuracy and stability of the drone during landing.

Introduction

Autonomous landing remains a significant challenge for rotor Unmanned Aerial Vehicles (UAVs), particularly when high precision is required for safe recovery in unpredictable environments. Traditional systems rely heavily on Global Positioning Systems (GPS) and Inertial Navigation Systems (INS). However, these can fail in "urban canyons," indoor environments, or areas with high electromagnetic interference.

While computer vision (detecting landing markers) and LiDAR (3D environmental mapping) offer potential solutions, they possess individual weaknesses: vision is sensitive to lighting and weather conditions, while LiDAR has a narrow field of view and high computational demands. This research proposes a robust autonomous landing system that bridges this "reliability gap" by creating a unified framework leveraging the strengths of multiple sensors simultaneously through Kalman filtering and Dempster-Shafer (DS) evidence theory.

Methodology

The system utilizes a multi-layered architecture for data acquisition and fusion, focusing on three primary components:

LiDAR Processing

The system uses LiDAR to scan the landing area to ensure a stable landing surface. The processing pipeline involves down-sampling via voxel grid filtering and denoising through smoothing filters. To create a complete 3D model from multiple angles, the system utilizes two primary registration algorithms: Iterative Closest Point (ICP) and Normal Distributions Transform (NDT).

The LiDAR data processing pipeline, from scanning to final registration.

Binocular Vision

Parallel binocular cameras are employed to detect a specific QR code landing marker. The system utilizes a dual-strategy approach based on distance:

  • Long-range: Employs OTSU thresholding and morphological operations to identify the QR code as a single object when details are blurry.
  • Short-range: Uses Canny edge detection to find "Position Detection Patterns" (the squares in the corners) for precise attitude and distance estimation.

Multi-Sensor Fusion Algorithm

The core innovation lies in a two-stage fusion process. First, individual sensor data is filtered via Kalman Filters. Second, Dempster-Shafer (DS) Evidence Theory calculates the "confidence" of each sensor. For example, if the camera is obscured by glare, its weight decreases, and the system relies more heavily on LiDAR or IMU data. A final Kalman Filter pass refines this weighted data to produce the final position estimate.

Key Insights and Ideas

  • Dynamic Weighting via DS Theory: Unlike basic fusion systems that assume sensors are always equally reliable, this system acknowledges that environmental factors change which sensor should be trusted most at any given millisecond.
  • Hierarchical Vision Pipeline: By treating QR code detection differently at long vs. short ranges, the system ensures a continuous "lock" on the landing target throughout the entire descent phase.
  • LiDAR as a Geometry Anchor: While vision provides the identity of the landing marker, LiDAR provides the structural context of the ground, ensuring the drone lands on a stable, flat surface.

Results and Discussion

The proposed architectural framework demonstrates improved robustness and precision. By integrating DS Evidence Theory, the system maintains stable positioning even when individual sensors experience high noise or temporary failure. Combining the high-frequency updates of an IMU with the absolute positioning of vision and the structural accuracy of LiDAR achieves the precision required for autonomous operations.

Limitations and Future Work

Despite its robustness, the system faces challenges regarding real-time performance. Running complex LiDAR registration (ICP/NDT) and binocular vision processing simultaneously requires significant onboard computational power. Future research will focus on optimizing these algorithms to reduce latency and exploring how deep learning might further enhance sensor configuration across a wider variety of complex application scenarios.

Summary updated

Abstract

With the rapid development of drone technology, the autonomous landing of drones has become a current research hotspot. This article introduces an autonomous drone landing system based on multi-sensor fusion. The system combines GPS, IMU, visual sensors, LiDAR, and other sensor data through Kalman filtering to achieve data fusion and improve the positioning accuracy and stability of the drone during landing.
BibTeX
@article{li2024research,
  author = {Jintang Li},
  title = {{Research on Precise Landing of Rotor UAV Based on Multi-Sensor Fusion}},
  journal = {International Journal of Computer Science and Information Technology},
  year = {2024},
  doi = {10.62051/ijcsit.v4n2.35},
}
Figure