svo_cg

Modified version of rpg_svo(commit d616106 on May 9, 2017) which implement a Semi-direct Monocular Visual Odometry pipeline

---

[TOC]

Dependencies

• Boost - C++ Librairies (thread and system are needed)

• Eigen 3 - Linear algebra

• OpenCV - Computer vision library for loading and displaying images

• Sophus - Lie groups

  • recommend version: commit id a621ff

• Fast - Corner Detector

• g2o - General Graph Optimization OPTIONAL

  • recommend version: commit id ff647bd (ff647bd7537860a2b53b3b774ea821a3170feb13)

    export G2O_ROOT=$HOME/installdir

• rpg_vikit (git submodule)

Download

git clone https://github.com/cggos/svo_cg.git --recursive

# or
git clone https://github.com/cggos/svo_cg.git
git submodule init
git submodule update

Build

with ROS

• catkin_make or catkin build

without ROS

• build vikit_common

  cd 3rdParty/rpg_vikit/vikit_common
  gedit CMakeLists.txt # SET(USE_ROS FALSE)
  mkdir build & cd build
  cmake .. & make -j3
  sudo make install

• build svo

  mkdir build & cd build
  cmake -DUSE_ROS=FALSE .. & make -j3

Run

with ROS

• on a Dataset (airground_rig_s3_2013-03-18_21-38-48.bag)

  roslaunch svo_ros test_rig3.launch [rviz:=true]
  rosbag play airground_rig_s3_2013-03-18_21-38-48.bag

• on a live camera stream (eg. RealSense ZR300 fisheye camera)

  # 1. calibrate your camera and modify `svo_ros/param/camera_atan_zr300.yaml`
  # 2. run
  roslaunch svo_ros realsense_zr300.launch  [rviz:=true]

without ROS

• on a Dataset (sin2_tex2_h1_v8_d.tar.gz)

  export SVO_DATASET_DIR=${HOME}/Datasets  # .bashrc
  
  source ~/.bashrc
  cd ${SVO_DATASET_DIR}
  [tsocks] wget http://rpg.ifi.uzh.ch/datasets/sin2_tex2_h1_v8_d.tar.gz -O - | tar -xz
  
  cd svo
  mkdir build & cd build
  cmake .. & make
  ./test_pipeline

Camera Calibration

Supported Camera Model

• ATAN model (preference)
  • which is also used by PTAM and uses the FOV distortion model
  • calibration tool: PTAM Calibration
• Pinhole model
  • three radial and two tangential distortion parameters
  • calibration tool: camera_calibration (ROS)
• Ocam model (by Davide Scaramuzza)
  • model cameras with high field of view or even omnidirectional cameras
  • calibration tool: OCamCalib toolbox

Import Calib Params

use the python scripts in scripts directory

• kalibr_to_svo.py
• omni_matlab_to_rpg.py

Notation

• px - Pixel coordinate (u,v)
• f - Bearing vector of unit length (x,y,z)
• T_f_w - Rigid body transformation from world frame w to camera frame f
• transforms a point in world coordinates p_w to a point in frame coordinates p_f as follows

  p_f = T_f_w * p_w

• camera position in world coordinates must be obtained by inversion

  pos = T_f_w.inverse().translation()

Obtaining Best Performance

In order to obtain the same performance as shown in the videos, consider the following:

• Use a global shutter camera (we use a grayscale matrix-vision bluefox camera, WVGA resolution).
• Set the framerate of the camera as high as possible (we use 70fps).
• Depending on your camera driver it might be better to manually fix the shutter speed and gain to avoid flickering.
• Use a lens with large field of view (we have approx. 110 deg).
• Calibrate the camera with the ATAN model and make sure you have a very low reprojection error (~0.1px).
• For higher robustness, you can increase the number of tracked features by setting the parameter svo/max_fts to 180.
• Avoid motions of pure rotation.
• The keyframe selection is currently designed for downlooking cameras. Forward motions are not performing well at the moment.

Notes

• Forward motions
  • The current keyframe selection criterion is designed for downward looking cameras. This is one reason why SVO does not work well for forward motions (e.g., to be used on a car).
• Image resolution
  • The current parameters are tuned for WVGA resolution. If you use a higher resolution camera, then the pyramid levels should be increased as well as the reprojection error thresholds.

Papers

@inproceedings{Forster2014ICRA,
  author = {Forster, Christian and Pizzoli, Matia and Scaramuzza, Davide},
  title = {{SVO}: Fast Semi-Direct Monocular Visual Odometry},
  booktitle = {IEEE International Conference on Robotics and Automation (ICRA)},
  year = {2014}
}

SVO 2.0

SVO 2.0: Fast Semi-Direct Visual Odometry for Monocular, Wide Angle, and Multi-camera Systems.

SVO 2.0 (IEEE TRO'17) extends the original SVO implementation (ICRA' 14) with the addition of edgletes, IMU prior, wide angle cameras (fisheye and catadioptric), multi-camera configurations, and forward looking camera motion.