sichitong / kinect_odom

Odometry from a Kinect sensor

kinect_odom

Odometry from a Kinect sensor

This is a project to play around with visual odometry using a kinect.

Current Status

The basic functionality is now in place, and the algorithm produces reasonable looking velocity measurements.

The algorithm (for now) is as follows:

1. Detect features in the previous frame
2. Track those features using Lucas & Kanade Optical Flow to get a corresponding set of features in the current frame
3. Project both feature sets into 3D using the kinect depth map and a pinhole camera model
4. Use the Iterative Closest Point algorithm to align the two 3D point clouds
5. Repeat!

Future Work

The end goal is to mimic the algorithm found in "Real-Time Stereo Visual Odometry for Autonomous Ground Vehicles" (Howard, 2008). This will require:

1. Use FAST to detect features
2. Match features using an n^2 SAD matcher, instead of LK tracking
3. Use Levenberg-Marquardt to solve the alignment, instead of ICP