Real-Time UAV Autonomous Localization Based on Smartphone Sensors

Boxin Zhao, Xiaolong Chen, Xiaolin Zhao, Jun Jiang, Jiahua Wei2018 · DOI

TL;DR

This paper presents a novel, real-time, and memory-efficient autonomous localization method for small UAVs using only smartphone sensors, overcoming the limitations of GPS-denied environments. By carefully evaluating and fusing data from smartphone accelerometers, gyroscopes, pressure sensors, and orientation sensors, the proposed approach achieves accurate localization without requiring complex environment mapping, making it suitable for resource-constrained UAVs.

What Problem Does This Paper Address?

Small Unmanned Aerial Vehicles (UAVs) hold immense potential for applications like rescue, exploration, and environmental monitoring. However, their reliance on GPS for navigation severely limits their operation in environments where GPS signals are unstable or unavailable, such as indoors, tunnels, or dense urban areas. While smartphones offer a compelling solution due to their high integration, light weight, and multi-sensor capabilities, their built-in sensors often suffer from low accuracy and noise. Furthermore, the limited storage and processing power of smartphones make it challenging to directly adapt traditional, resource-intensive localization algorithms like Visual Odometry (VO) or Visual Simultaneous Localization and Mapping (VSLAM). This paper tackles this bottleneck by exploring the feasibility of using smartphone sensors for accurate UAV localization and developing a lightweight, real-time algorithm tailored for these constraints.

Methods

The core of this work lies in a two-pronged approach: sensor evaluation and algorithm development.

  1. Smartphone Sensor Evaluation: The authors meticulously evaluated the accuracy and reliability of key smartphone sensors:
  2. Gyroscope and Accelerometer: They characterized bias stability and repeatability over time and developed calibration methods to estimate installation error matrices. Experiments involved controlled rotations on a turntable and motion on an air track, comparing sensor readings against ground truth from high-precision instruments.
  3. Orientation Sensor: This software sensor, integrating accelerometer and geomagnetic data, was evaluated for its pitch and roll accuracy using a turntable. A lookup table was proposed to calibrate its measurements.
  4. Pressure Sensor: Its accuracy for altitude estimation was assessed by measuring descent in a tall building, demonstrating a reliable measurement of relative height.
  5. Multi-Phone Sensors Fusion-Based Visual Odometry (MFVO): This algorithm builds upon existing visual odometry frameworks but is specifically adapted for smartphone constraints.
  6. Key-frame based approach: It focuses on tracking features and estimating camera poses relative to key frames rather than building a full environment map, significantly reducing memory requirements.
  7. Sensor Fusion: Crucially, it fuses visual information with the calibrated smartphone orientation sensor and pressure sensor measurements to estimate camera motion scale and ultimately, the UAV's position.
  8. No Environment Mapping: This design choice is central to its efficiency and suitability for resource-limited platforms.

Datasets

The paper does not explicitly mention using pre-existing public datasets for training. Instead, it relies on:

  • Custom Experimental Data: The authors conducted extensive experiments using a Samsung Galaxy S3 smartphone mounted on various platforms (turntable, air track, Parrot AR drone, XAircraft 650 quadrotor). Ground truth was obtained through high-precision instruments (e.g., photoelectric instruments, digital timers, airborne GPS) or established algorithms like AprilTag.
  • Comparison with Existing Algorithms: The proposed MFVO algorithm was compared against TUM_PTAM [28] and a previous method (MLVO [33]) using data collected during their experiments.

Key Insights and Ideas

  • Smartphone Sensors are Viable for UAV Localization: Despite their inherent noise and lower accuracy compared to dedicated navigation sensors, smartphone MEMS sensors, when properly calibrated and fused, can provide sufficient data for real-time UAV localization.
  • Bias Instability Requires Pre-Calibration: The study highlights that MEMS accelerometer and gyroscope biases are not perfectly stable, especially during initial operation. This necessitates a pre-calibration and filtering step before each deployment for reliable performance.
  • Pressure and Orientation Sensors Offer Good Precision: The pressure sensor demonstrated reliable relative height measurements, and the orientation sensor, with calibration, provided good pitch and roll estimates, which are crucial for attitude estimation.
  • Map-less Visual Odometry is Key for Efficiency: By avoiding the need to build and maintain an environment map, the MFVO algorithm significantly reduces computational and memory overhead, making it practical for smartphones.
  • Fusion Enhances Robustness: Combining visual data with inertial and pressure sensor readings helps to overcome the limitations of individual sensors and improve overall localization accuracy and robustness.

Key Results

  • Sensor Accuracy:
    • Gyroscope: Mean error around 0.001-0.003 rad/s after calibration.
  • Accelerometer: Mean error around 0.05-0.15 m/s² after calibration, with 98% of points on the fitting curve for X/Y axes and 92% for the Z-axis.
  • Pressure Sensor: Average measurement error of 0.28 m with a standard deviation of 0.026 m for relative height estimation.
  • Algorithm Performance (Indoor - Parrot AR Drone):
    • MFVO achieved a localization RMSE of 0.42 m.
  • This was significantly better than TUM_PTAM's RMSE of 0.71 m.
  • The MFVO path closely matched the AprilTag ground truth.
  • Algorithm Performance (Outdoor - Wheat Fields):
    • MFVO achieved an RMSE of 2.28 m in X and 4.54 m in Y, an improvement of 30.41% over MLVO.
  • Localization error increased with distance, approximately 8% of the flight distance.
  • Algorithm Performance (Outdoor - Urban Environment):
    • The algorithm showed good initial tracking but drifted due to wind and moving vehicles.
  • Localization RMSE was around 8.1 m after 200 m, approximately 4.05% of the flight distance.

Limitations and Future Work

  • Bias Instability: The paper acknowledges that MEMS sensor biases are unstable initially and require pre-calibration. While they propose calibration, continuous online calibration or more advanced filtering could further improve robustness.
  • Sensitivity to Dynamic Environments: The urban experiment highlighted that moving objects (cars) can disrupt visual tracking, leading to localization failures. This is a significant limitation for real-world deployment in complex environments.
  • Yaw Angle Vulnerability: The orientation sensor's yaw angle was found to be easily affected by magnetic interference, making it unreliable in certain environments.
  • Future Work:
    • Improving Accuracy in Dynamic Environments: Developing strategies to handle moving objects and dynamic scenes is a key area for future research.
  • Computational Time and Resource Consumption: While the algorithm is designed to be efficient, further analysis and optimization of computation time and resource usage will be considered.
  • Control Strategy Development: Investigating how to integrate the proposed localization algorithm with UAV control systems for autonomous flight is a crucial next step.

Summary updated

BibTeX
@misc{zhao2018real,
  author = {Boxin Zhao and Xiaolong Chen and Xiaolin Zhao and Jun Jiang and Jiahua Wei},
  title = {{Real-Time UAV Autonomous Localization Based on Smartphone Sensors}},
  year = {2018},
  doi = {10.3390/s18124161},
}
Figure