libing64 / monocular_vo

Monocular Visual Odometry

1.1 Steps

• Map Init

  • feature detection
  • feature tracking
  • motion estimation wtih 2d-2d correspondence
  • triangulation

• Pose Update

  • feature detection with mask
  • feature tracking
  • motion estimation with 3d-2d correspondence
  • traingulation for new detected features
  • local bundle adjustment
  • update triangulated points to map

1.2 Building

cd catkin_ws/src
git clone [e-mail]:libing64/mono_vo.git
cd ..
catkin_make -DCATKIN_WHITELIST_PACKAGES="mono_vo"

1.3 Running with kitti dataset

modify the dataset_folder in mono_vo_kitti.launch

soure devel/setup.bash
roslaunch mono_vo mono_vo_kitti.launch

1.5 Test Environment

Ubuntu 20.04 + ros noetic

2. TODO

2.1 How to make the estimator more robust and accurate?

• local bundle adjustment
• record screen to gif
• mapping
• fusing imu
• sliding window
• clean outliers

3. Stereo-VO vs Mono-VO

单目VO比双目VO难在哪？ 双目VO通过单帧的一对图像可以直接获取深度信息， 前后两帧之间可以通过3d-2d或者3d-3d的方法估计运动，然后进行积分。

但是单目没法直接获取深度信息，前后帧之间虽然可以估计的R和t，但是平移信息只有方向没有尺度，需要sliding window保存多帧的信息，通过多帧之间的约束关系才能保证尺度一致。