SJTU-ViSYS/Ground-Fusion

What it solves

Ground-Fusion addresses the challenge of reliable localization and mapping for ground vehicles using low-cost sensors. It specifically targets "corner cases"—situations where individual sensors fail or degrade, such as visual challenges (poor lighting or textureless areas), wheel odometer slips, or GNSS signal loss/interference—which often cause standard SLAM systems to fail.

How it works

The system uses a factor graph to tightly integrate data from four different sensor sources: RGB-D images, inertial measurements (IMU), wheel odometry, and GNSS signals. To maintain robustness, it employs three key strategies:

1. Adaptive Initialization: It uses three distinct initialization methods (stationary, visual, and dynamic) to ensure the system starts correctly regardless of the initial state of the vehicle.
2. Anomaly Detection: It includes mechanisms to detect when a specific sensor is providing degraded or erroneous data and handles these anomalies to prevent them from corrupting the localization estimate.
3. Dense Mapping: It provides real-time dense color mapping of the environment.

Who it’s for

This project is designed for robotics researchers and developers building ground vehicles that need accurate, real-time localization and mapping in diverse indoor and outdoor environments without relying on expensive lidar sensors.

Highlights

• Lidar-less SLAM: Achieves state-of-the-art performance without the need for expensive lidar hardware.
• Multi-Sensor Fusion: Tightly integrates RGB-D, IMU, wheel odometry, and GNSS.
• Robustness to Corner Cases: Specifically designed to handle sensor degradation and anomalies.
• Comprehensive Benchmarks: Accompanied by open-source datasets (M2DGR, M2DGR-plus, and Ground-Challenge) for validation.