HITSZ-NRSL/Dynamic-VINS

What it solves

Dynamic-VINS addresses the challenge of maintaining accurate robot localization (odometry) in dynamic environments where moving objects can confuse standard visual-inertial systems. It is specifically designed to run in real-time on resource-restricted edge platforms, such as the NVIDIA Jetson AGX Xavier and HUAWEI Atlas200 DK.

How it works

The system uses a combination of RGB-D cameras and inertial sensors (IMU). It identifies and removes dynamic features (moving objects) using a combination of object detection and depth information, achieving performance similar to semantic segmentation but with lower overhead. To maintain efficiency, it employs grid-based feature detection and high-quality FAST feature extraction.

Who it’s for

This project is for robotics developers and researchers working on autonomous navigation and visual-inertial odometry (VIO) for robots with limited onboard computing power.

Highlights

• Dynamic Object Handling: Uses object detection and depth data to recognize and filter out dynamic features.
• Resource Efficient: Optimized for real-time performance on edge devices like Jetson AGX Xavier and Atlas200 DK.
• Robust Localization: Maintains competitive accuracy and robustness even in environments with moving people or objects.
• Multi-Platform Support: Provides specific configurations and launch files for PyTorch, TensorRT, and Huawei Ascend hardware.