UAVD4L: A Large-Scale Dataset for UAV 6-DoF Localization
(2401.05971)Abstract
Despite significant progress in global localization of Unmanned Aerial Vehicles (UAVs) in GPS-denied environments, existing methods remain constrained by the availability of datasets. Current datasets often focus on small-scale scenes and lack viewpoint variability, accurate ground truth (GT) pose, and UAV build-in sensor data. To address these limitations, we introduce a large-scale 6-DoF UAV dataset for localization (UAVD4L) and develop a two-stage 6-DoF localization pipeline (UAVLoc), which consists of offline synthetic data generation and online visual localization. Additionally, based on the 6-DoF estimator, we design a hierarchical system for tracking ground target in 3D space. Experimental results on the new dataset demonstrate the effectiveness of the proposed approach. Code and dataset are available at https://github.com/RingoWRW/UAVD4L
Overview
-
UAVD4L addresses the challenge of UAV localization without GNSS systems by providing a comprehensive dataset for research.
-
The dataset enhances traditional limitations by adding textured 3D models and synthetic data for accurate 6-DoF localization.
-
UAVLoc, a two-stage pipeline, combines synthetic data generation and visual localization with IMU data integration for accurate pose estimation.
-
A novel hierarchical target tracking system improves object localization by converting 2D image coordinates to 3D map locations.
-
The dataset promises advances in UAV technology for GPS-denied areas and includes diverse scenes and ground truth data, with future work focusing on varied lighting conditions.
Introduction to UAVD4L
The reliance on GNSS systems for Unmanned Aerial Vehicles (UAVs) localization in environments where GPS signals are unavailable presents a challenge for various applications. Traditional datasets available for UAV localization are often limited in scope and accuracy, inhibiting progress in GPS-denied areas. A new dataset, UAVD4L, develops a comprehensive solution to overcome these challenges.
Tackling Limitations in Existing Datasets
UAVD4L overcomes the limitations of small-scale and limited viewpoint datasets by incorporating a textured 3D reference model and synthetic data such as rendered RGB and depth images. This approach allows for greater variability in viewpoints and more accurate 6-DoF ground truth poses without relying on potentially inaccurate built-in sensor data.
The Two-Stage 6-DoF Localization Pipeline
The UAVD4L dataset introduces a two-stage localization pipeline named UAVLoc. The first stage involves offline synthetic data generation to cover various viewpoints and conditions. The second stage conducts an online visual localization process where rotation information from the UAV's own Inertial Measurement Unit (IMU) narrows the search for relevant reference images. Next, feature matches establish 2D-3D correspondences which enable a gravity-guided PnP RANSAC to estimate the UAV's camera pose accurately.
Hierarchical 3D Target Tracking System
Apart from the dataset, a novel hierarchical target tracking system has been designed to capitalize on 6-DoF localization results. The system uses two lenses to pinpoint the exact location of objects on the ground, projecting their position from 2D images to absolute coordinates on a 3D map. This system could significantly enhance tracking accuracy in applications such as surveillance and search-and-rescue operations.
Conclusion and Future Work
UAVD4L's large-scale dataset for UAV 6-DoF localization represents an important step in advancing UAV technologies in GPS-denied environments. It provides a valuable benchmark for research, offering a dataset that includes a diverse range of urban and rural scenes, and accurate ground truth pose estimations. While the UAVD4L has achieved considerable accuracy, future work aims to include more challenging lighting conditions to extend the dataset's applicability. The provision of open-source code and data ensures that researchers and practitioners can continue to build on this promising foundation.
Create an account to read this summary for free:
GitHub - RingoWRW/UAVD4L (45 stars)