• A robust visual-inertial state estimator algorithm for dynamic environments.
• The robust bundle adjustment is used for dynamic objects such as cars 🚗, humans 🏃, buses 🚌, etc.

• Please refer our paper for detailed explanantions and experimental results!
  • The BA module validated on VIODE dataset dataset.
  • The loop closure module validated on Our dataset.
  • Algorithm is based on VINS-Fusion.
• Youtube video for the detailed explanation and the demo video.

• In our paper, VINS-Fusion combined with Switchable Constraints (git)was compared with DynaVINS.
• VINS-Fusion-SC, an algorithm that integrates Switchable Constraints into the loop closure module of VINS-Fusion, is also useful, so we released the algorithm.
• You can visit here to use VINS-Fusion-SC.
• When using VINS-Fusion-SC for your paper, don't forget to cite our paper!:wink:

This code is tested on

• Linux 18.04 LTS
• ROS Melodic
• Ceres Solver 1.14.0
• OpenCV 3.4.1

The dependency of DynaVINS is equal to that of VINS-Fusion.

Ubuntu 64-bit 16.04 or 18.04. ROS Kinetic or Melodic. ROS Installation

Follow Ceres Installation.

Due to the limiting file size of Github, we need one package and two files from the VINS-Fusion repository.

1. Set the camera_models package in your workspace, which is included in VINS-Fusion.
2. Copy support_files/brief_k10L6.bin in VINS-Fusion into our support_files folder
3. Copy support_files/brief_pattern.yml in VINS-Fusion into our support_files folder

Please follow the below commands to build DynaVINS (on ROS).

$ cd ~/catkin_ws/src 
$ git clone https://github.com/url-kaist/dynaVINS
$ cd ../
$ catkin_make  
(or if you use catkin tools) catkin build
$ source ~/catkin_ws/devel/setup.bash

For convenience, we also provide 3_high.bag file in the parking_lot scene. You can download the file by the following command:

$ wget https://urserver.kaist.ac.kr/publicdata/dynaVINS/VIODE_dataset/parking_lot/3_high.bag

• Since it is blocked by network provider, download the dataset from : http://gofile.me/4355j/XIv6xyQ2i

Note that the larger the number of bag files in the VIODE dataset is, the more dynamic objects exist.

$ roslaunch dynaVINS viode_mono.launch
$ rosbag play 3_high.bag (or 0_none.bag, 1_low.bag, ...)

$ roslaunch dynaVINS viode_stereo.launch
$ rosbag play 3_high.bag (or 0_none.bag, 1_low.bag, ...)

You can use your own intel realsense d455! (calibration is required)

You can easily download our bag file by the following command:

$ wget https://urserver.kaist.ac.kr/publicdata/dynaVINS/d455_urban_robotics/e_shape.bag

• Since it is blocked by network provider, download the dataset from : http://gofile.me/4355j/XIv6xyQ2i

$ roslaunch dynaVINS d455_mono.launch
$ rosbag play e_shape.bag (or loop_tempstatic.bag, ...)

$ roslaunch dynaVINS d455_stereo.launch
$ rosbag play e_shape.bag (or loop_tempstatic.bag, ...)

Parameters of DynaVINS. You can find the results of each parameter on the wiki page (param)

Time comparison according to various parameters can be found on the wiki page (time).

• regularization_lambda

  The Lambda value of regularization term in paper. (Section III-C)

• momentum_on

  Using the momentum factor or not (true/false)

• momentum_lambda

  The Lambda value of momentum term in paper. (Section III-D)

• alternating_converge

  The threshold for checking the convergence of the alternating optimization.
  90% is usually enough. If you want faster speed, please try to reduce it.
  Time comparison can be found on the wiki page.

• margin_feature_thresh

  Features which have less weight than this value are not used in marginalization.
  This may affect accuracy, but is effective at reducing time costs.
  You can try uncomment line 848 of "vins_estimator/estimator/estimator.cpp" to enable these features also in optimization.

  //ADDITIONAL FEATURE : NO USE IN OPT.
  //if(it_per_id.weight<DYN_FEAT_MARGIN_THRESH) continue;

• hypodiff_dist

  The distance threshold for grouping constraints into the hypothesis. (Section IV-B)

• hypodiff_yaw

  The angle threshold for grouping constraints into the hypothesis. (Section IV-B)

• hypo_regularization

  The Lambda value of regularization term in loop closure module. (Section IV-C)

• hypo_alternating_converge

  The threshold for checking the convergence of the alternating optimization in the loop closure module.

If you use our codes, please cite our paper (arXiv)

@article{song2022dynavins,
    title={DynaVINS: A Visual-Inertial SLAM for Dynamic Environments},
    author={Song, Seungwon and Lim, Hyungtae and Lee, Alex Junho and Myung, Hyun},
    journal={IEEE Robotics and Automation Letters},
    year={2022}
}

If you have any questions, please do not hesitate to contact us:

• Seungwon Song ✉️ sswan55 at kaist.ac.kr
• Hyungtae Lim ✉️ shapelim at kaist.ac.kr