Visual Odometry with a Single-Camera Stereo Omnidirectional System

Proof of concept of the visual odometry (VO) application via a single-camera catadioptric stereo omnidirectional system (SOS) and performance comparison against a popular RGB-D camera.

Author: Carlos Jaramillo
Contact: [email protected]

Copyright (C) 2019 under the Gnu Public License version 3 (GPL3)

WARNING: This code is at an experimental stage, and has only been tested under Python 3.6 on Mac OS X 10.12.6 (Sierra)

Data Sets

The data sets associated with this project can be downloaded from the Datasets Section.

Software Requirements

Python3 is required, with the following modules:

Numpy
Matplotlib
Vispy - PyQt5
Scipy (Optional, for GUMS calibration)

For instance Using Python 3.6, with pip3, install the following modules:

$ pip3 install numpy
$ pip3 install scipy
$ pip3 install sympy

For visualization:

$ pip3 install matplotlib
$ pip3 install vispy
$ pip3 install pyqt5

In addition, we require these packages with their respective Python3 bindings:

OpenCV
OpenGV

See the end of this README for how to install OpenCV and my custom forked of OpenGV

Running the Demos

Navigate to the path where this project was cloned. Then, invoke python3 with the following usages.

To run the VO demo with the RGB-D camera:

usage: demo_vo_rgbd.py [-h] [--is_synthetic IS_SYNTHETIC]
                       [--hand_eye_transformation HAND_EYE_TRANSFORMATION]
                       sequence_path

Demo of frame-to-frame visual odometry for the RGB-D images collected for the
vo_single_camera_sos project.

positional arguments:
  sequence_path         The path to the sequence where the rgbd folders is
                        located.

required arguments:
  --is_synthetic IS_SYNTHETIC
                        Determines whether the data is real or synthetic. This
                        is necessary to use the appropriate camera intrinsic
                        parameters.

optional arguments:
  --hand_eye_transformation HAND_EYE_TRANSFORMATION
                        (Optional) If real-life data, this indicates the
                        complete path and name to hand-eye transformation file
                        
  --visualize_VO VISUALIZE_VO
                        (Optional) Indicates whether to visualize the
                        estimated 3D trajectory (and ground-truth if
                        available).

For example, using the free_style sequence,

$ python3 demo_vo_rgbd.py "PATH_TO_MY_SEQUENCE/free_style" --is_synthetic=false --hand_eye_transformation="PATH_TO_APPROXIMATED/rgbd_hand_eye_transformation.txt" --visualize_VO=true

To run the VO demo with the single-camera SOS:

usage: demo_vo_sos.py [-h] [--calibrated_gums_file CALIBRATED_GUMS_FILE]
                      sequence_path

Demo of frame-to-frame visual odometry for the Single-camera SOS images
collected for the vo_single_camera_sos project.

positional arguments:
  sequence_path         The path to the sequence where the omni folder is
                        located.

required arguments:
  --calibrated_gums_file CALIBRATED_GUMS_FILE
                        Indicates the complete path and name of the calibrated
                        GUMS pickle file
                        
optional arguments:
  --visualize_VO VISUALIZE_VO
                        (Optional) Indicates whether to visualize the
                        estimated 3D trajectory (and ground-truth if
                        available).

For example, using the free_style sequence,

$ python3 demo_vo_sos.py "PATH_TO_MY_SEQUENCE/free_style" --calibrated_gums_file="PATH_TO_CALIBRATED_MODEL/gums-calibrated.pkl" --visualize_VO=true

Detailed Setup for OpenCV and OpenGV

OpenCV 3

The following guide uses Homebrew for Mac OS X:

Requirements:

Make sure you have installed the XCode CLI tools are all installed:

$ ls -lah /usr/include/sys/types.h

If not, try:

$ xcode-select --install

The last release of OpenCV's Deep Neural Network module requires Google's Protocol Buffers

$ brew install protobuf

However, it may fail to find a "suitable threading library available." If so, disable DNN within the CMake configuration (in the next steps)

Install OpenCV 3

The Easy Way Via Homebrew (Mac OS X only):

$ brew install opencv3

The Hard (more powerful) Way from Source Code:

Clone OpenCV 3 from the repo

$ cd ~/src
$ git clone [email protected]:Itseez/opencv.git

And also the contributed modules:

$ git clone [email protected]:Itseez/opencv_contrib.git

Configure the installation vis CMake (or Curses CMake as I'd like to do):

$ cd opencv/..
$ mkdir opencv-build
$ cd opencv-build
$ ccmake ../opencv -G"Eclipse CDT4 - Unix Makefiles"

The extra modules path is set like :

OPENCV_EXTRA_MODULES_PATH        /PATH_TO_/src/opencv_contrib/modules

For example, Python paths should be set to:

 PYTHON2_EXECUTABLE               /usr/local/bin/python2.7
 PYTHON2_INCLUDE_DIR              /usr/local/Cellar/python/2.7.14/Frameworks/Python.framework/Versions/Current/include/python2.7
 PYTHON2_INCLUDE_DIR2
 PYTHON2_LIBRARY                  /usr/local/Cellar/python/2.7.14/Frameworks/Python.framework/Versions/Current/lib/libpython2.7.dylib
 PYTHON2_LIBRARY_DEBUG
 PYTHON2_NUMPY_INCLUDE_DIRS       /usr/local/lib/python2.7/site-packages/numpy/core/include
 PYTHON2_PACKAGES_PATH            lib/python2.7/site-packages
 PYTHON3_EXECUTABLE               /usr/local/bin/python3
 PYTHON3_INCLUDE_DIR              /usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/include/python3.6m
 PYTHON3_INCLUDE_DIR2
 PYTHON3_LIBRARY                  /usr/local/Cellar/python3/3.6.3/Frameworks/Python.framework/Versions/3.6/lib/libpython3.6m.dylib
 PYTHON3_LIBRARY_DEBUG
 PYTHON3_NUMPY_INCLUDE_DIRS       /usr/local/lib/python3.6/site-packages/numpy/core/include
 PYTHON3_PACKAGES_PATH            /Users/carlos/src/opencv-build/lib/python3.6/site-packages

Then, compile OpenCV 3, and install as usual:

$ make
$ make install

In Ubuntu Linux

In Ubuntu 16.04, using the ros-kinetic-opencv3 version and workaround to use it with python3:

Install OpenCV3 for ROS:

$ sudo apt-get install ros-kinetic-opencv3
Install python3 versions of rospkg and catkin

$ sudo -H pip3 install rospkg catkin_pkg

NOTE for Python3: Once the usual ros-kinetic-opencv3 package is installed and working for python2.7, you can get it working in python3 by removing the catkin sourced in the .bashrc file. Otherwise, just adjust your IDE not to have the ROS python paths for the $PYTHONPATH.

In other words:

Problems are caused by ROS adding /opt/ros/kinetic/lib/python2.7/dist-packages to the $PYTHONPATH. This actually happens when you activate ROS with the command source /opt/ros/kinetic/setup.bash. This line is often added at the end of your .bashrc file, in /home/username/.bashrc.
A workaround is to remove this line from the bashrc file. This way the python3 opencv packages will be correctly used, and you can still run source /opt/ros/kinetic/setup.bash to use ROS. However, this does mean you cannot use ROS and python3 from the same environment.

You need to install the following module via pip3:

$ sudo -H pip3 install opencv-python
Test OpenCV under python3:

$ python3

import cv2

OpenGV (for Visual Odometry):

pyopengv depends on Boost >= 1.66 and Boost.python.numpy, so make sure it gets installed before proceeding.

In Mac OS X, with Homebrew:

$ brew install boost
$ brew install boost-python3

In Ubuntu, you may need to compile boost 1.66+ by downloading its source as indicated in this tutorial. I did:

$ cd path/to/boost_1_67_0
$ ./bootstrap.sh --help

Select your configuration options and invoke ./bootstrap.sh again, for example:

$ ./bootstrap.sh --with-python=python3

Finally,

$ sudo ./b2 install

Compile and install OpenGV

Assuming, Boost >= 1.66 and *Boost.python.numpy were already successfully installed.

Clone the fork from opengv because this has Python bindings for the non-central camera case, which we need for tracking the omnistereo system pose:

$ cd ~/src
$ git clone https://github.com/ubuntuslave/opengv.git
$ cd opengv
$ git checkout non_central-python
$ cd .. && mkdir opengv-build && cd opengv-build

Now build the project, making sure

The correct paths for Python get selected, and
The boost.python.numeric path is set correctly.

$ ccmake ../opengv

If compiling with python3 support, you must toggle the advanced configuration and set the appropriate Python paths.

For example, in Mac OS X, I set:

Boost_NUMPY3_LIBRARY_DEBUG       /usr/local/lib/libboost_numpy36-mt.dylib                        
Boost_NUMPY3_LIBRARY_RELEASE     /usr/local/lib/libboost_numpy36-mt.dylib                        
Boost_NUMPY_LIBRARY_DEBUG        /usr/local/lib/libboost_numpy-mt.dylib                          
Boost_NUMPY_LIBRARY_RELEASE      /usr/local/lib/libboost_numpy-mt.dylib

PYTHON_EXECUTABLE             /usr/local/bin/python3
PYTHON_INCLUDE_DIR            /usr/local/Cellar/python3/3.6.5/Frameworks/Python.framework/Versions/3.6/include/python3.6m
PYTHON_INSTALL_DIR            /usr/local/lib/python3.6/site-packages/
PYTHON_LIBRARY                /usr/local/Cellar/python3/3.6.5/Frameworks/Python.framework/Versions/3.6/lib/libpython3.6m.dylib

For example, in Ubuntu 16.04, I had:

PYTHON_EXECUTABLE             /usr/bin/python3
PYTHON_INCLUDE_DIR            /usr/include/python3.5
PYTHON_INSTALL_DIR            /usr/local/lib/python3.5/dist-packages
PYTHON_LIBRARY                /usr/lib/x86_64-linux-gnu/libpython3.5m.so

In Ubuntu, I had to also configure the proper location of the installed-from-source Boost 1.66:

Boost_DIR                        /home/carlos/src/boost_1_66_0                                                  
Boost_INCLUDE_DIR                /usr/local/include                                                             
Boost_LIBRARY_DIR_DEBUG          /usr/local/lib                                                                 
Boost_LIBRARY_DIR_RELEASE        /usr/local/lib                                                                 
Boost_PYTHON3_LIBRARY_DEBUG      /usr/local/lib/libboost_python3.so                                             
Boost_PYTHON3_LIBRARY_RELEASE    /usr/local/lib/libboost_python3.so                                             
Boost_PYTHON_LIBRARY_DEBUG       /usr/lib/x86_64-linux-gnu/libboost_python.so                                   
Boost_PYTHON_LIBRARY_RELEASE     /usr/lib/x86_64-linux-gnu/libboost_python.so

Once the Cmakefiles are generated, compile and install as usual:

$ make
$ sudo make install
$ sudo ldconfig

NOTE: I had to run $ sudo ldconfig for the $ python3 -c "import pyopengv" to work.

Name		Name	Last commit message	Last commit date
Latest commit History 28 Commits
omnistereo		omnistereo
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
demo_vo_rgbd.py		demo_vo_rgbd.py
demo_vo_sos.py		demo_vo_sos.py
demo_vo_sos_live.py		demo_vo_sos_live.py
setup.py		setup.py

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

Visual Odometry with a Single-Camera Stereo Omnidirectional System

Data Sets

Software Requirements

Running the Demos

To run the VO demo with the RGB-D camera:

To run the VO demo with the single-camera SOS:

Detailed Setup for OpenCV and OpenGV

OpenCV 3

Requirements:

Install OpenCV 3

The Easy Way Via Homebrew (Mac OS X only):

The Hard (more powerful) Way from Source Code:

In Ubuntu Linux

OpenGV (for Visual Odometry):

Compile and install OpenGV

About

Releases

Packages

Languages

License

ubuntuslave/vo_single_camera_sos

Folders and files

Latest commit

History

Repository files navigation

Visual Odometry with a Single-Camera Stereo Omnidirectional System

Data Sets

Software Requirements

Running the Demos

To run the VO demo with the RGB-D camera:

To run the VO demo with the single-camera SOS:

Detailed Setup for OpenCV and OpenGV

OpenCV 3

Requirements:

Install OpenCV 3

The Easy Way Via Homebrew (Mac OS X only):

The Hard (more powerful) Way from Source Code:

In Ubuntu Linux

OpenGV (for Visual Odometry):

Compile and install OpenGV

About

Resources

License

Stars

Watchers

Forks

Releases

Packages 0

Languages

Packages