3D Computer Vision · Stereo Visual SLAM

Stereo visual odometry and SLAM in Python, built around feature tracking, multi-view geometry, local bundle adjustment, pose-graph optimization, and loop closure. Experiments use KITTI Odometry sequence 00 to evaluate camera trajectories, reprojection errors, feature consistency, and pose uncertainty.

Stack: Python · OpenCV · NumPy · GTSAM · Matplotlib
Areas: Visual SLAM, geometric computer vision, nonlinear optimization, graph-based inference

Estimated camera trajectories compared with KITTI reference poses.

Pipeline

flowchart LR
    A[Rectified stereo frames] --> B[Feature detection and matching]
    B --> C[Stereo triangulation]
    C --> D[Temporal tracking and PnP / RANSAC]
    D --> E[Feature track database]
    E --> F[Local bundle adjustment]
    F --> G[Pose graph]
    G --> H[Loop verification and global optimization]

Loading

The system progresses from frame-to-frame pose estimation to globally consistent trajectory refinement. Loop candidates are filtered using geometric and uncertainty-based checks before additional constraints enter the pose graph.

Implementation

Stage	Methods	Code
Stereo data and calibration	KITTI stereo image loading, projection matrices, reference poses	dataset_utils.py
Feature association	AKAZE, stereo and temporal correspondence checks	feature_tracking.py
Geometry and pose	Triangulation, reprojection error, PnP/RANSAC	geometry.py
Track management	Indexed observations, track histories, serialization	tracking_database.py
Local optimization	Keyframes, GTSAM bundle adjustment	bundle_adjustment.py
Global optimization	Relative pose factors, pose-graph updates	pose_graph.py
Loop closure	Candidate filtering, Mahalanobis checks, verification	loop_closure.py
Measurement and plotting	Trajectory error, tracking statistics, uncertainty analysis	performance_analysis.py, visualization.py

Experiment drivers ex1.py–ex7.py cover the stages in sequence, from pairwise geometry and feature association to loop-closure experiments. They are not a single command-line pipeline; later stages depend on intermediates produced earlier.

Evaluation on KITTI 00

The final evaluation reports the following for an experiment covering 3,300 frames:

Metric	Value
Feature tracks recorded	438,490
Mean track length	4.74 frames
Mean feature links per frame	629.69

Feature tracking and inliers	Pose uncertainty

The results compare incremental PnP estimates, bundle-adjusted poses, and pose graphs with/without loop closure. Detailed methodology, plots, and interpretation are in the technical report. Figures describe the original experimental configuration and are not a KITTI leaderboard submission.

Setup and execution

Dataset: KITTI Odometry, grayscale stereo sequence 00, camera calibration, and reference poses. Dataset files are not included. The exact layout is documented in Dataset setup.

Python dependencies: OpenCV, NumPy, GTSAM, and Matplotlib (see requirements.txt). A compatible Python/GTSAM combination is required.

python -m venv .venv
.\.venv\Scripts\Activate.ps1
python -m pip install -r requirements.txt
cd code
python ex1.py

Run from code/, because the implementation uses relative paths into ../dataset/. The tracking stage (ex4.py) generates a serialized feature database required by downstream stages. See the exercise scripts and the dataset setup guide for data requirements.

The complete dataset-dependent workflow has not been re-run with a clean installation; dependency versions are not pinned to a tested environment.

Technical skills

Stereo geometry and calibration · robust model fitting · feature tracking · nonlinear least squares · factor graphs · uncertainty modeling · loop-closure verification · trajectory evaluation · scientific visualization.

Documentation

Individual experiment reports are available under docs/reports/, alongside the final analysis. The datasets and large serialized tracking artifacts are intentionally external to the repository.