VINS-GNC

This package implements Visual-Inertial Navigation System (VINS) applied to mobile robots working in dynamic environment. The implementation is based on VINS-Fusion. Also, it's enhanced by wheel odometry based on VIW-Fusion.

Original paper: Robust Visual-Inertial Odometry for Ground Robots in Dynamic Environments.

1. Installation

Create a catkin workspace where the package will be built. In order to simplify deployment process, Dockerfile is provided. Build an image and run the container:

cd docker/ubuntu
sudo usermod -aG docker $USER
bash docker_build.sh
xhost +local:docker
bash docker_run.sh /path/to/workspace

Inside a docker container, run the following commands to build catkin workspace:

conda init
source .bashrc
conda activate vins_env

source /opt/ros/noetic/setup.bash

catkin init
catkin config --extend $(rospack find roscpp | sed 's|/share/roscpp$||')
catkin config --merge-devel
catkin config -DCMAKE_BUILD_TYPE=Release -DCATKIN_PYTHON_EXECUTABLE=$(which python)

cd src
git clone https://github.com/WFram/Robust-VINS.git
cd $HOME

catkin build vins
source devel/setup.bash

2. Run

2.1. VIODE dataset

Run the following commands:

# 1st shell (visualizer)

roslaunch vins vins_rviz.launch

# 2st shell (node)
roslaunch vins viode_vio.launch

# 3rd shell (bag)
rosbag play /path/to/bag

2.2. Run with semantic segmentation

First, NN model needs to be provided. Can be downloaded from mmsegmentation storage.

wget -O /path/to/config/nn_model.pth https://download.openmmlab.com/mmsegmentation/v0.5/segformer/segformer_mit-b0_512x512_160k_ade20k/segformer_mit-b0_512x512_160k_ade20k_20210726_101530-8ffa8fda.pth

Second, parameter semantics in YAML config file has to be set. Finally, VINS node must be run with the additional argument:

roslaunch vins <LAUNCH_FILE> semantics:=true

3. Parameter description

• camera

  Use visual constraints in optimization (turn off only for inertial integration debug)

• semantics

  Use semantic information provided

• imu

  Use inertial integration and constraints in optimization

• no_inertial_constraints

  Use IMU data only for pose prediction. No inertial constraints added into optimization. Not recommended to set

• wheel

  Use wheel odometry readings provided (for wheeled mobile robots)

• no_wheel_constraints

  Use wheel odometry data only for pose prediction. No wheel constraints added into optimization. Not recommended to set

• plane

  Use plane constraint based on planar motion. Reduced accumulated drift along vertical axis

• number_of_cameras

  1: mono setup 2: stereo setup

• semantic_palette

  Dataset which was used to train a network, whose checkpoints are provided. The only possible options are: ade20k, cityscapes, pascal_voc

• predict_by_wheels

  If true, then instead of inertial data, wheel odometry measurements are used to predict initial pose for optimization. Required if visual-wheel setup used

• resize

  Whether to downscale images by factor 2. NOTE: in order to use it, camera calibration intrinsic provided in YAML files have to be resized manually

• dilating_kernel_size

  Kernel size for semantic masks dilation. Useful for avoiding dynamic points on object borders to be used in optimization

• roll_n

  Roll planar constraint noise

• pitch_n

  Pitch planar constraint noise

• zpw_n

  Z-axis planar constraint noise. Decreasing can provide better reducing drift over vertical motion axis

• max_cnt

  The number of features maintaned by feature extractor in an image stream

• min_dist

  The minimal distance between extracted features on an image

• use_ransac

  Use outlier rejection based on fundamental matrix and RANSAC. Useful for mono setups. In some cases, may diverge initialization performance

• F_threshold

  Threshold for outlier rejection based on RANSAC (px)

• max_depth

  Landmarks with larger depth values are filtered out

• flow_back

  Use optical flow check for improving tracking results. May decrease runtime

• gnc

  Use dynamic landmark rejection based on Graduated Non-Convexity. Additional sensor (IMU, wheel encoders) is required

• loss

  Loss function used for dynamic landmarks rejection in optimization. Options: L1: 0; Geman-McClure: 1; Leclerc (recommended): 2

• fixed_shape

  Constant shape parameter of a loss function. Larger values increase sensitivity to outliers

• convexity_init

  Initial value of convexity. Has to be larger than convexity_margin. Larger values increases inference time of optimization, but increases outlier rejection performance

• convexity_margin

  Margin convexity value to reach before finishing alternating optimization. Not recommended to change (set to 1 by default)

• convexity_update

  Constant for updating shape of a loss function. Larger values decrease runtime, but lead to less stable optimization process. Not recommended to change (set to 1.4 by default)

• feature_weight_threshold

  Landmarks with weigths less than given will not participate in optimization

• regularizer

  Constant factor to regularize loss function. Not recommended to change (set to 2 by default)

• squared_eps

  Multiplier for reprojection errors. Not recommended to change (set to 1 by default)

• max_solver_time

  Optimization solver maximal time. Larger values lead to better accuracy but larger inference time

• max_num_iterations

  Maximal solver inner interation number. Larger values lead to better accuracy but larger inference time

• keyframe_parallax

  Parallax threshold for selecting keyframes (pixel). Smaller values lead to more keyframes, that increases runtime, but improves accuracy. Might deteriorate performance in dynamic environment

• acc_n

  IMU accelerometer noise density

• gyr_n

  IMU gyro noise density

• acc_w

  IMU accelerometer random walk

• gyr_w

  IMU gyro random walk

• estimate_td_inertial

  Online calibration of timeshift between camera and IMU

• td_inertial

  Prior timeshift value between camera and IMU