luohongk/SuperVINS

📖[IEEE Sensors Journal (JSEN) ] SuperVINS: A Real-Time Visual-Inertial SLAM Framework for Challenging Imaging Conditions (integrated deep learning features)

What it solves

SuperVINS addresses the failure of traditional Visual-Inertial SLAM (Simultaneous Localization and Mapping) systems in challenging imaging conditions. It replaces handcrafted feature extraction and matching—which often struggle with lighting changes or low texture—with deep-learning-based methods to improve robustness and accuracy in real-time.

How it works

Built on the VINS-Fusion backbone, SuperVINS integrates deep learning models via ONNX Runtime GPU to handle the entire visual pipeline:

  • Feature Extraction: Uses SuperPoint to detect learned keypoints and descriptors instead of Shi-Tomasi corners.
  • Feature Matching: Employs LightGlue for learned matching instead of traditional optical flow (KLT).
  • Loop Detection: Utilizes DBoW3 combined with SuperPoint descriptors.
  • Loop Verification: Combines SuperPoint and LightGlue matching with PnP RANSAC to verify loop closures more reliably.

Who it’s for

This framework is designed for robotics researchers and developers building autonomous systems that need reliable localization and mapping in environments where traditional visual features are unstable.

Highlights

  • Deep Learning Integration: Fully replaces traditional front-end and loop closure modules with SuperPoint and LightGlue.
  • Real-Time Performance: Optimized for GPU execution using ONNX Runtime.
  • Comprehensive Output: Provides real-time VIO trajectories, pose-graph optimized trajectories, and dense loop-corrected trajectories in TUM format.
  • Docker Support: Includes a pre-built Docker image for simplified setup of CUDA and ROS Noetic dependencies.

Related

  • Project
    openvinotoolkit/openvino

    An open-source toolkit for optimizing and deploying deep learning models across CPU, GPU, and NPU hardware to boost inference performance.

  • Project
    luigifreda/pyslam

    A hybrid Python/C++ modular framework for Visual SLAM that integrates modern AI depth prediction, semantic segmentation, and volumetric reconstruction for research and prototyping.

  • Project
    PeterFWS/Structure-PLP-SLAM

    A visual SLAM system that uses points, lines, and planes for efficient sparse mapping and localization across monocular, RGB-D, and stereo cameras.

  • Project
    sjtuyinjie/Ultra-Fusion

    A resilient, tightly-coupled multi-sensor SLAM framework that maintains accurate localization for wheeled, legged, and aerial robots even during sensor degradation or calibration errors.

  • Project
    cvg/glue-factory

    A library for training and evaluating deep neural networks that extract and match local visual features, supporting state-of-the-art models like LightGlue and GlueStick.