/visual_odom

This repository is C++ OpenCV implementation of Stereo Odometry

Primary LanguageC++MIT LicenseMIT

Stereo Visual Odometry

example workflow

This repository is C++ OpenCV implementation of Stereo Visual Odometry, using OpenCV calcOpticalFlowPyrLK for feature tracking.

Reference Paper: https://lamor.fer.hr/images/50020776/Cvisic2017.pdf

Demo video: https://www.youtube.com/watch?v=Z3S5J_BHQVw&t=17s

alt text

alt text

Requirements

OpenCV 3.0
If you are not using CUDA:

sudo apt update
sudo apt install libopencv-dev 

If you use CUDA, compile and install CUDA enabled OPENCV. check InstallOPENCV.md

Dataset

Tested on KITTI odometry dataset.

Compile & Run

git clone https://github.com/ZhenghaoFei/visual_odom.git

The system use Camera Parameters in calibration/xx.yaml, put your own camera parameters in the same format and pass the path when you run.

mkdir build
cd build
cmake ..
make -j4
./run /PathToKITTI/sequences/00/ ../calibration/kitti00.yaml

GPU CUDA acceleration

Thanks to temburuyk, the most time consumtion function circularMatching() can be accelerated using CUDA and greately improve the performance. 60~80 FPS on a decent NVIDIA Card. To enable GPU acceleration

  1. Make sure you have CUDA compatible GPU.
  2. Install CUDA, compile and install CUDA supported OpenCV
  3. When compiling, use
cmake .. -DUSE_CUDA=on
  1. Compile & Run

Reference code

  1. Monocular visual odometry algorithm

  2. Matlab implementation of SOFT