Monocular Visual SLAM (WIP)

Monocular visual odometry — the front end of a SLAM system — estimating camera motion and 3D structure from a single camera. Backend components like loop closure and global optimization are not covered here.

Takes single-camera frames from the KITTI odometry benchmark and:

• Detects and matches features (ORB/SIFT) between consecutive frames
• Estimates relative pose via Essential Matrix (5-point algorithm + USAC_MAGSAC RANSAC)
• Accumulates frame-to-frame transforms into a global trajectory (SE(3))
• Triangulates a sparse, color-sampled point cloud from matched features
• Visualizes trajectory and point cloud in Open3D with ground truth comparison

Translation is recovered only up to an unknown scale factor — the fundamental limitation of monocular SLAM.

Tech stack: Python, OpenCV, Open3D, NumPy, pykitti, spatialmath

Project Structure

File	Description
slam.py	Two-class pipeline: Processor (preprocessing, ORB/SIFT, FLANN/BF matching, essential matrix → R,t) and VisualOdometry (main loop, SE(3) accumulation, triangulation, color sampling from cam2)
kitti.py	Wraps pykitti.odometry — frame iterator over cam0 (grayscale) and cam2 (RGB), projection matrices, ground truth poses
utils.py	Matching helpers (ratio test, cross-check), coordinate transforms, keypoint distribution (SSC, adaptive grid), blur detection
visualization.py	Map class — Open3D GUI on main thread, receives SLAM updates via post_to_main_thread(). Estimated path (blue) vs. ground truth (green)
performance.py	FPS / timing monitor
config.py	Paths, target resolution, debug flags

Usage

1. Download the KITTI Odometry Dataset

Download from the KITTI odometry page:

• Grayscale sequences (required for odometry)
• Color sequences (required for point cloud coloring)
• Ground truth poses
• Calibration files

Expected folder structure (pykitti convention):

dataset/
└── sequences/
    ├── 00/
    │   ├── image_0/   # cam0 — grayscale left
    │   ├── image_2/   # cam2 — color left
    │   └── calib.txt
    ├── 01/
    │   └── ...
    └── ...
    poses/
    ├── 00.txt
    ├── 01.txt
    └── ...

2. Configure and run

1. Set BASE_PATH (path to dataset/) and KITTI_SEQUENCE in config.py
2. Tune DETECTOR, MATCH_METHOD, and other parameters as needed
3. Run:

python slam.py

Dependencies

numpy
opencv-python
open3d
imutils
spatialmath-python
pykitti
Pillow

Install with:

pip install -r requirements.txt