This is a simple visual odometry pipeline that estimates the relative pose between two monocular frames using a fundamental matrix. The goal is for this pipeline to develop full-fledged monocular SLAM capabilities with support for sensors such as inertial measurement units and global positioning.
Blue line is the output from the pipeline, green line is ground truth trajectory.
I build the code against the latest versions of OpenCV, Ceres solver and glog that's available in archlinux. I use pangolin v0.6 due to a bug in the later versions thatI haven't gotten around to fixing, but I will look into it soon.
- Clone the repo.
mkdir build && cd build && cmake .. && make main- Hope you are lucky.
The build emits the binary main, to which you pass in the following
arguments.
-
video file
This is any valid string that can be passed to OpenCV's VideoCapture. To use a directory of images, use OpenCV's globbing style, like so:
/mnt/datasets/kitti-odometry/seq0/dataset/sequences/00/image_0/%6d.png -
frame_id
How many frames to process before stopping
-
f
Focal length in pixel units
-
cx cy
Image center coordinates in the image
-
n_features
Unused, can be used to set the number of features to detect before matching
-
poses_file
A file containing ground truth poses formatted like the KITTI ground truths. It is optional. If provided, estimated translations are scaled to match the data and ground truth trajectory is shown in the Pangolin display.