Visual Odometry for CPET Dataset

This is my final project for ROB501: Computer Vision for Robotics. Here I create a simple visual odometry pipeline based on Large-Scale Direct Stereo SLAM (LSD-SLAM)[1] to localize a rover on Canadian Planetary Emulation Terrain (CPET) dataset. There is no mapping, simply localization, and thus the localization estimate tends to drift significantly. However, it is a proof of concept for semi-dense depth map based visual odometry using stereo cameras.

Functionality

Using stereo images from one of the rover's runs across the Martian-like terrain, disparity images are generated using the get_disparity.py function, a disparity map generator based on census transforms [2]. Using non-linear optimization, the homogenous transformation from one position to the next are calculated. The UTM position of the rover at each increment is found using these transforms and plotted onto an overhead image of the terrain.

Results

Read my final report in Dense Visual Odometry.pdf.

Install

1. git clone this repository and create an input and an output folder beside the src folder. Call them input and output.
2. Download raster_data, run1_base_hr/omni_image4, and run1_base_hr/omni_image5 from the CPET dataset above and put the folders in the input folder. Do not change the name of the files or folders.

Your folder structure should look like:

input/
    run1_base_hr/
        omni_image4/
        omni_image5/
    raster_data/
output/
src/

3. Install docker desktop

Run

1. Simply run docker-compose up (or sudo docker-compose up) which will pull the correct docker container and run the code. This will generate the disparity images of all images from start_frame to end_frame and plot the path on the overhead image. You can change these two variables in rob501_project.py.