Visual Odometry Pipeline

Conda environment

Install conda and than run this:

conda env create --file vision.yml

Run the Pipeline

# usage: main.py [-h] [--parking] [--kitty] [--malaga] [--plot] [--plot2]

# optional arguments:
#   -h, --help     show this help message and exit
#   --parking, -p  the parking dataset
#   --kitty, -k    the kitty dataset
#   --malaga, -m   the malaga dataset
#   --plot, -o     plot the results
#   --plot2, -o2   plot ground truth vs vo pipeline

# Example for plotting and selecting the parking dataset
# plot2 plots the groundtruth and vo pipeline output (only the camera poses (only kitti and parking))
python src/main.py -o -p

Download datasets

To download the zipfiles for the project you can use this script. (I hope it works on windows/mac)

python data/download.py

Tasks

• how to use dataclass => code example
• put init camera pose stuff into a init file
• Add PnP and write code example with the just the next frame
• add visualisation that shows which 3D points the next camera posed used
• implemented continuous pipeline (works with all frames)
• add feature matches visualisation (use heatmap if image is not working)
• KLT Tracker: Use this algorithm to track the pixels inside the matches

Bonus

• Boundle Adjustment
• Record Dataset
• Loop detection
• Compare Sift vs Harris vs FAST vs SURF
• (Kalman Filter)

Done

• Feature detection and matching ( Use library, OpenCV, use efficent version?, SIFT / RANSAC ? ) $\sqrt{}$
• Setup pipeline and visualization $\sqrt{}$
• Triangulation $\sqrt{}$

Documentation

Major classes and important architecture decisions