Related Experiment Videos
A Pose Initialization Method for Unmanned Vehicles Based on an Improved Siamese Neural Network and Multi-Stage
Jian Yang1, Biao Chen1, Weiye Shen1
1College of Mechanical Engineering, Yangzhou University, Yangzhou 225127, China.
Sensors (Basel, Switzerland)
|June 12, 2026
Summary
This study introduces a novel pose initialization framework for satellite-denied environments. It uses a Siamese Neural Network (SNN) and multi-stage registration to achieve centimeter-level accuracy for autonomous systems.
Area of Science:
- Robotics and Autonomous Systems
- Computer Vision
- Sensor Fusion
Background:
- Conventional localization methods fail in satellite-denied environments due to lack of GPS.
- Rapid and precise pose initialization is critical for autonomous navigation in GPS-denied areas.
Purpose of the Study:
- To develop a high-precision pose initialization framework for satellite-denied environments.
- To enable accurate localization using radar point clouds and prior maps.
Main Methods:
- A Siamese Neural Network (SNN) with CBAM for feature matching between radar point clouds and map slices.
- Adaptive Monte Carlo Localization (AMCL) for probabilistic slice identification and initial pose refinement.
- Normal Distributions Transform (NDT) for centimeter-level pose estimation.
Main Results:
- The SNN successfully identifies correct map slices, with AMCL and NDT refining pose accuracy to centimeter-level.
- Achieved 99% localization success rates with low RMSEs for distance and orientation on various map scales.
- Demonstrated robust initialization performance on the KITTI dataset in complex outdoor environments.
Conclusions:
- The proposed multi-stage framework offers a reliable solution for high-precision pose initialization in large-scale satellite-denied scenarios.
- The integration of SNN, AMCL, and NDT significantly enhances localization accuracy and robustness.
