Visual SLAM tutorial

A visual SLAM (simultaneous localization and mapping) framework provides the foundation for integrating different components within it. For example, A visual SLAM system comprises camera tracking, mapping, loop closing via place recognition, and visualization components. The framework connects the components such that we get the camera motion and the structure of the environment from a stream of images in real-time.

Objective
Framework
Setup
Build
Run
Demo
Contributing

Objective

The main goal of this project is to create an ease-of-use framework to learn visual SLAM while allowing for customizability for better performance. To this end, we define the components in the framework as plugins, which can be modified so long as the interfaces (i.e., the inputs and outputs) are unchanged. Academic researchers could also use it to write new plugins to verify hypotheses.

Framework

We organize the visual SLAM components as individual plugins/libraries that can be tested before integrating them into the framework. The plugins are loaded into ROS2 composable nodes to run them in a single process. The nodes and plugins are described as follows:

data loader node takes a list of images from a folder or an image stream from a camera (using a camera plugin) and converts them into the Frame message, which contains the image and the camera information.
vslam node processes the incoming Frame messages to track the camera motion and map the structure (represented by points). The system begins in the initialization state, where we set the current frame as the tentative keyframe and attempt initialization in the next frame. Any failure in tracking and mapping resets the state back to initialization. Once the system is initialized, we track the camera pose of the images against the keyframe and create new keyframe and map points (structure of the scene) as needed. In a parallel thread, we check if the camera revisits a previously mapped area and optimize the camera motion and structure globally. In the event of failure to find correspondences between two subsequent frames, the system enters the relocalization state, where we attempt to regain camera tracking. This node consists of seven plugins:
- feature extraction plugin calculates the keypoints (e.g., corners or high-gradient image regions) and, optionally, their descriptors in the image.
- feature matcher plugin finds feature correspondences between the keypoints in the current frame and the nearest keyframe in the back-end during the tracking state.
- camera tracker plugin calculates the relative pose between the current frame and the current keyframe based on the feature correspondences.
- mapper plugin triangulates a set of new map points based on the relative pose and the feature correspondences.
- place recognition plugin finds the old keyframe image similar to the image in the current frame.
- back-end plugin runs local bundle adjustment and pose-graph optimizations.
- visualizer plugin gets the Frame messages and update the visualizer accordingly.

Setup

The code has been tested on Ubuntu and macOS running ROS 2 Humble
- Installation instructions for Ubuntu
- Installation instructions for macOS
Download KITTI odometry dataset (color, 65 GB)

Clone this repository in your home directory

cd ~/
git clone https://github.com/yan99033/VisualSLAMTutorial.git

Build

Go to the root of the project directory (cd ~/VisualSLAMTutorial)
Run colcon build --cmake-args -DCMAKE_BUILD_TYPE=Release to build the packages.

Run

Create a copy of the KITTI camera parameters (and remove the .example extension) and modify the parameters accordingly.
Create a copy of the vslam_demo parameters (and remove the .example extension) and modify the parameters accordingly.

Launch the demo

cd ~/VisualSLAMTutorial
source install/setup.bash
ros2 launch vslam_demos vslam_from_folder.launch.py params:=<path_to_test_kitti_yaml_file>

Replace the first command with source install/setup.zsh if you are using macOS.

Demo

KITTI 00

Contributing

Thanks for considering contributing to the project. The guidelines are as follows:

If you encounter a bug
- If it has not been reported
  - Create a new issue, describe the issue and steps to reproduce the bug (this may include the dataset and parameter settings).
- Else
  - Check the existing issues and see if they help.
  - Create a PR to resolve an existing issue
Else if you request a new feature
- Use the issue tracker to discuss the new feature.

fgxhxx/VisualSLAMTutorial