ACFR-RPG/DynOSAM

Offical code release for DynoSAM: Dynamic Object Smoothing And Mapping. Accepted Transactions on Robotics (Visual SLAM SI). A visual SLAM framework and pipeline for Dynamic environements, estimating for the motion/pose of objects and their structure, as well as the camera odometry and static map.

What it solves

DynoSAM addresses the challenge of Dynamic SLAM (Simultaneous Localization and Mapping), where a robot must estimate its own position and map its environment while simultaneously tracking the motion of moving objects. Traditional SLAM often struggles in dynamic environments because moving objects can confuse the camera's pose estimation and lead to inaccurate maps.

How it works

DynoSAM uses a Stereo or RGB-D Visual Odometry pipeline to estimate camera poses, object motions, and the static background. It employs a modular architecture with a front-end for data processing and a back-end for optimization.

Key technical components include:

  • Optimization Procedures: Supports full-batch, sliding-window, and incremental optimization.
  • Object Detection: Integrates YOLOv8 for instance segmentation and tracking, accelerated by CUDA and TensorRT.
  • Tracking: Uses either dense optical flow (RAFT) or sparse KLT tracking.
  • Backend Formulations: Offers multiple motion estimation modes, including World-Centric Motion Estimator (WCME), World-Centric Pose Estimator (WCPE), and Hybrid formulations.
  • Integration: Fully integrated with ROS2 for data handling and RVIZ for 3D visualization.

Who it’s for

This project is designed for robotics researchers and engineers working on visual SLAM, autonomous navigation in dynamic environments, and embodied intelligence.

Highlights

  • Hardware Acceleration: CUDA and TensorRT integration for faster front-end processing and object detection.
  • Edge Device Support: Tested on NVIDIA ORIN NX (AARM64).
  • Flexible Backends: Multiple optimization formulations to balance speed and accuracy (e.g., Parallel Hybrid for speed).
  • Comprehensive Evaluation: Includes built-in scripts for reproducing paper results and generating detailed metrics (ATE, RPE).
  • IMU Support: Integration with GTSAM for IMU preintegration.

Related

  • Project
    nvidia-isaac/cuVSLAM

    A CUDA-accelerated library for visual odometry and SLAM that provides real-time, high-accuracy spatial tracking across various camera and IMU configurations without requiring parameter tuning.

  • 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
    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
    stella-cv/stella_vslam

    stella_vslam is an open‑source visual SLAM library (monocular, stereo, RGB‑D) that supports many camera models, lets you save/load maps, and offers modular APIs plus ROS integration. It’s a community fork of OpenVSLAM, actively maintained under a BSD‑2 license.

  • Project
    MIT-SPARK/VGGT-SLAM

    A real-time dense feed-forward scene reconstruction system for 3D mapping and localization, featuring optional open-set 3D object detection.