This is an OpenCV 3.0 based implementation of a monocular visual odometry algorithm.
##Algorithm Uses Nister's Five Point Algorithm for Essential Matrix estimation, and FAST features, with a KLT tracker. More details are available here as a report, and here as a blog post.
Note that this project is not yet capable of doing reliable relative scale estimation, so the scale informaion is extracted from the KITTI dataset ground truth files.
##Demo Video
##Requirements OpenCV 3.0
##How to compile? Provided with this repo is a CMakeLists.txt file, which you can use to directly compile the code as follows:
mkdir build
cd build
cmake ..
make##How to run? After compilation, in the build directly, type the following:
./vo##Before you run In order to run this algorithm, you need to have either your own data, or else the sequences from KITTI's Visual Odometry Dataset. In order to run this algorithm on your own data, you must modify the intrinsic calibration parameters in the code.
##Performance
##Use own datas
##Something wrong of the code First I use two fixed images, then I find the roll-pitch-yaw are strange. (It looks quite different from my own code) Then I put the camera above a car to record a series of frames, it's roughly right. But when the car stop, it looks strange (I think it's because of scale). And the translation looks strange, may be because the rotation is wrong.