Monocular Visual Odometry Pipeline - Vision Algorithms for Autonomous Robots

This repository contains a locally consistent monocular visual odometry pipeline. This pipeline is based on the lecture contents of the course "Vision Algorithms for Autonomous Robots" at ETH Zurich/UZH taught by Prof. Dr. Davide Scaramuzza in the spring autumn semester 2024/25. The pipeline estimates the camera position (vehicle pose) with respect to the camera frame of the first image from a sequence of images. The pipeline was tested on three different datasets, which can be accessed on the Robot Perception Group website. The datasets are: KITTI, Parking and Malaga.

Dependencies

The pipeline was implemented in Python 3.8 and the required packages are given in the conda environment file environment.yml. To create the conda environment, run the following command:

conda env create -f cvenv.yml

and activate the environment with:

conda activate cvenv

Using a virtual environment, the required packages are given in the requirements.txt file. To create the virtual environment, run the following command:

pip install -r requirements.txt

Datasets

The datasets for testing the pipeline should be stored in the data directory. The datasets can be downloaded from the Robot Perception Group website. The datasets are: KITTI, Parking and Malaga. The datasets should be stored in the following directory structure:

data/
    kitti/
        ...
    parking/
        ...
    malaga/
        ...

Running the pipeline

To run the pipeline, execute the following command:

python main.py

The pipeline will run on the KITTI dataset by default. To run the pipeline on the Parking or Malaga dataset, change the dataset index in the main.py file.

ds = 0  # 0: KITTI, 1: Malaga, 2: parking

The pipeline will output the estimated camera poses as a trajectory plot with the ground truth trajectory, extracted features and the corresponding 3D points. The plot will be saved in the 'out' directory.

out/interface_plot.png

For a continuous plotting of the estimated trajectory the according flag in the main.py file can be set.

interface_plot = True

Structure:

The VO-pipeline in 'VisualOdometry.py' is structured in:

• Initialization: Bootstrapping the pipeline with two images
  • Initial feature detection using sift
  • Initial feature matching using brute force matching and ratio test
  • Initial pose estimation using essential matrix for known intrinsics
• Continuous Operation: Estimating the camera pose for each frame
  • Feature detection
  • Feature tracking using KLT (assuming sufficiently small frame to frame motion)
  • Triangulation of 3D points from two views
  • Pose estimation using PnP RANSAC from 3D-2D correspondences