VID-Fusion

VID-Fusion: Robust Visual-Inertial-Dynamics Odometry for Accurate External Force Estimation

Authors: Ziming Ding , Tiankai Yang, Kunyi Zhang, Chao Xu, and Fei Gao from the ZJU FAST Lab.

0. Overview

VID-Fusion is a work to estimate odometry and external force simultaneously by a tightly coupled Visual-Inertial-Dynamics state estimator for multirotors. Just like VIMO, we formulate a new factor in the optimization-based visual-inertial odometry system VINS-Mono. But we compare the dynamics model with the imu measurements to observe the external force and formulate the external force preintegration like imu preintegration. So, the thrust and external force can be added into the classical VIO system such as VINS-Mono as a new factor.

We present:

• An external force preintegration term for back-end optimization.
• A complete, robust, tightly-coupled Visual-Inertial-Dynamics state estimator.
• Demonstration of robust and accurate external force and pose estimation.

Simultaneously estimating the external force and odometry within a sliding window.

Related Paper: VID-Fusion: Robust Visual-Inertial-Dynamics Odometry for Accurate External Force Estimation, Ziming Ding, Tiankai Yang, Kunyi Zhang, Chao Xu, and Fei Gao, ICRA 2021.

Video Links: bilibili or Youtube.

1. Prerequisites

Our software is developed and only tested in Ubuntu 16.04, ROS Kinetic (ROS Installation), OpenCV 3.3.1.

Ceres Solver (Ceres Installation) is needed.

2. Build on ROS

cd your_catkin_ws/src
git clone [email protected]:ZJU-FAST-Lab/VID-Fusion.git
cd ..
catkin_make  --pkg quadrotor_msgs  # pre-build msg
catkin_make

3. Run in vid-dataset

cd your_catkin_ws
source ~/catkin_ws/devel/setup.bash
roslaunch vid_estimator vid_realworld.launch
roslaunch benchmark_publisher publish.launch #(option)
rosbag play YOUR_PATH_TO_DATASET

We provide the experiment data for testing, in which the vid-experiment-dataset is in ros bag type. The dataset provides two kinds of scenarios: tarj8_with_gt and line_with_force_gt.

• tarj8_with_gt is a dataset with odometry groundtruth. The drone flys with a payload.

• line_with_force_gt is a dataset with external force groundtruth. The drone connects a force sensor via a elastic rope.

A new visual-inertial-dynamics dataset with richer scenarios is provided in VID-Dataset.

The drone information should be provided in VID-Fusion/config/experiments/drone.yaml. It is noticed that you should use the proper parameter of the drone such as the mass and the thrust_coefficient, according to the related bag file.

As for the benchmark comparison, we naively edit the benchmark_publisher from VINS-Mono to compare the estimated path, and add a external force visualization about the estimated force and the ground truth force. The ground truth data is in VID-Fusion/benchmark_publisher/data. You should switch path or force comparison by cur_kind in publish.launch (0 for path comparison and 1 for force comparison).

As for model identification, we collect the hovering data for identification. For the two data bags, tarj8_with_gt and line_with_force_gt, we also provide the hovering data for thrust_coefficient identification. After system identification, you should copy the thrust_coefficient result to VID-Fusion/config/experiments/drone.yaml

roslaunch system_identification system_identify.launch 
rosbag play YOUR_PATH_TO_DATASET
#copy the thrust_coefficient result to VID-Fusion/config/experiments/drone.yaml

The external force is the resultant force except for rotor thrust and aircraft gravity. You can set force_wo_rotor_drag as 1 in config file to subtract the rotor drag force from the estimated force. And the related drag coefficient k_d_x and k_d_y should be given.

4. Acknowledgements

We replace the model preintegration and dynamics factor from VIMO, and formulate the proposed dynamics and external force factor atop the source code of VIMO and VINS-Mono. The ceres solver is used for back-end non-linear optimization, and DBoW2 for loop detection, and a generic camera model. The monocular initialization, online extrinsic calibration, failure detection and recovery, loop detection, and global pose graph optimization, map merge, pose graph reuse, online temporal calibration, rolling shutter support are also from VINS-Mono.

5. Licence

The source code is released under GPLv3 license.

6. Maintaince

For any technical issues, please contact Ziming Ding ([email protected]) or Fei GAO ([email protected]).

For commercial inquiries, please contact Fei GAO ([email protected]).