Efficient Global Indoor Localization for Micro Aerial Vehicles

V. Strobel, Roland Meertens, Guido C. H. E. de CroonarXiv 2016 · OpenReview

Abstract

Indoor localization for autonomous micro aerial vehicles (MAVs) requires specific localization techniques, since the Global Positioning System (GPS) is usually not available. We present an efficient onboard computer vision approach that estimates 2D positions of an MAV in real-time. This global localization system does not suffer from error accumulation over time and uses a $k$-Nearest Neighbors ($k$-NN) algorithm to predict positions based on textons—small characteristic image patches that capture the texture of an environment. A particle filter aggregates the estimates and resolves positional ambiguities. To predict the performance of the approach in a given setting, we developed an evaluation technique that compares environments and identifies critical areas within them. We conducted flight tests to demonstrate the applicability of our approach. The algorithm has a localization accuracy of approximately 0.6 m on a 5 m $\times$ 5 m area at a runtime of 32 ms on board of an MAV. Based on random sampling, its computational effort is scalable to different platforms, trading off speed and accuracy.

Introduction

Autonomous Micro Aerial Vehicles (MAVs) face a unique challenge in indoor environments. They are typically too small to carry heavy sensors like LiDAR, and GPS signals are unavailable inside buildings. While sophisticated visual SLAM (Simultaneous Localization and Mapping) algorithms exist, they are often too computationally expensive for the limited onboard processors of lightweight drones. Furthermore, local methods like optical flow suffer from "drift," where small errors accumulate over time. This research presents a "global" localization method that determines the drone's absolute position without a starting reference point, designed to be efficient enough for high-frame-rate execution on standard MAV hardware.

Methodology

The proposed framework shifts heavy computation to an offline phase, allowing for real-time performance during flight. The system pipeline consists of four primary stages:

  1. Texton Histograms: The system uses "textons"—small $6 \times 6$ pixel patches—to capture local texture. A dictionary of these patches is created offline. During flight, the drone extracts patches from its downward-facing camera to generate a histogram of texton frequencies.
  2. $k$-Nearest Neighbors ($k$-NN): This histogram is compared against a pre-recorded database of histograms with known $x,y$ coordinates. The $k$-NN algorithm identifies the most similar matches to provide initial position estimates.
  3. Particle Filter: To resolve positional ambiguities (e.g., different parts of a room looking similar), a particle filter tracks multiple hypotheses over time, filtering noise to provide a stable estimate.
  4. Map Evaluation: A novel "global loss" function identifies "critical areas" in an environment where textures are too repetitive or sparse for reliable localization.
System architecture overview showing the transition from raw image capture to texton feature extraction, k-NN position estimation, and particle filter aggregation.

Performance and Scalability

A key advantage of this approach is its scalability via random sub-sampling. By sampling only a fraction of the image patches to create the texton histogram, the system can trade off accuracy for speed. This allows the algorithm to be tuned for different hardware capabilities.

Performance trade-offs as a function of the number of samples, showing localization error and processing frequency.

In experimental trials over a $5 \times 5$ meter area, the system achieved a localization accuracy of approximately 0.6 meters. The total processing time per frame was just 32ms (roughly 31 Hz), which is significantly faster than traditional homography-based methods. In practical "triggered landing" tests, the MAV successfully identified target zones and landed within 16cm of the desired spot in most trials.

Discussion and Future Work

While the system provides robust global localization without drift, it currently assumes a relatively constant flight altitude and is sensitive to yaw (rotation). Future iterations could incorporate IMU data to "de-rotate" images or utilize multi-scale texton dictionaries to handle altitude changes. The most promising direction involves a hybrid approach: fusing this texton-based global localization with local optical flow to combine high precision with zero long-term drift.

Summary updated

Abstract

Indoor localization for autonomous micro aerial vehicles (MAVs) requires specific localization techniques, since the Global Positioning System (GPS) is usually not available. We present an efficient onboard computer vision approach that estimates 2D positions of an MAV in real-time. This global localization system does not suffer from error accumulation over time and uses a $k$-Nearest Neighbors ($k$-NN) algorithm to predict positions based on textons---small characteristic image patches that capture the texture of an environment. A particle filter aggregates the estimates and resolves positional ambiguities. To predict the performance of the approach in a given setting, we developed an evaluation technique that compares environments and identifies critical areas within them. We conducted flight tests to demonstrate the applicability of our approach. The algorithm has a localization accuracy of approximately 0.6 m on a 5 m$\times$5 m area at a runtime of 32 ms on board of an MAV. Based on random sampling, its computational effort is scalable to different platforms, trading off speed and accuracy.
BibTeX
@misc{strobel2016efficient,
  author = {V. Strobel and Roland Meertens and Guido C. H. E. de Croon},
  title = {{Efficient Global Indoor Localization for Micro Aerial Vehicles}},
  year = {2016},
  eprint = {1610.07233},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1610.07233},
}
Figure