Stereo Visual Odometry

This project is an implementation of a Stereo Visual Odometry (SVO) pipeline using KITTI Dataset. SVO is a method used in computer vision and robotics to estimate the 3D pose (position and orientation) of a camera relative to its starting position, using only the images captured by the camera.

Features

• Feature detection using SIFT (Scale-Invariant Feature Transform)
• Feature matching using BFMatcher (Brute-Force Matcher)
• 3D-2D motion estimation using PnP (Perspective-n-Point) with RANSAC (Random Sample Consensus)
• Triangulation of points using stereo camera parameters

Requirements

• Python 3.6 or higher
• OpenCV 4.0 or higher
• NumPy
• Pytorch

Usage

1. Clone the repository: git clone https://github.com/AdeolaJoseph/Stereo-Visual-Odometry.git

2. Navigate to the project directory: cd Stereo-Visual-Odometry

3. Run the main script: python stereo_odometry.py

4. Run eval scripts

• Copy and paste the result into the eval/result folder
• Navigate to the eval folder
• Run python compute_metrics.py

Please see devnotes.md for an explanation of current files, bugs, evaluation scripts etc.

Authors

1. Joseph Adeola
2. Khawaja Alamdar
3. Moses Ebere
4. Nada Abbas

License

This project is licensed under the MIT License - see the LICENSE file for details.

Results

APE vs RPE

The following figure shows the APE error on sequence 01. APE tries to fit the estimated trajectory over the ground truth and then computes the metrics. This is not ideal as the drift is minimal in the beginning and higher in the end, and APE punishes the trajectory in the beginning. It does not compute local trajectory errors.

Estimated Trajectory and Groundtruth Trajectory

Absolute Pose Error

Also, it will be sensitive to outlier poses as it is trying to force the two trajectories to align. Should the odometry be strictly penalized if after a couple of turns it had a few-degree error but after that performed flawlessly? [Ivan]

RPE, on the other hand, captures local trajectory errors.

For evaluation with others, it is better to use APE as it is standardized.

Compiled Results:

Sequence	APE - rmse	APE - rmse (%)
00	46.71	1.25
01	35.56	1.45
02	141.11	2.7
03	9.15	1.53
04	NA	NA
05	12.02	0.55
06	NA	NA
07	6.2	0.89
08	13.27	0.41
09	16.82	1.11
10	9.07	0.65

00

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

01

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

02

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

- APE

03

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

04

• APE

• RPE

05

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

06

• APE

• RPE

07

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

08

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

09

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error

10

Estimated Trajectory and Groundtruth Trajectory
Absolute Pose Error	Absolute Pose Error
Relative Pose Error	Relative Pose Error