Dynamic Region Detection for Indoor RGB-D Scenes

indoSLAM

This script detects dynamic regions between two RGB-D frames using depth-based clustering and ORB feature matching.

Requirements

• Python 3
• OpenCV
• NumPy
• scikit-learn

Install Dependencies

pip install opencv-python numpy scikit-learn

Folder Structure The script expects RGB and depth images organized as follows:

tests/
└── test1/
    ├── rgb/
    │   ├── 1.png
    │   └── 2.png
    └── depth/
        ├── 1.png
        └── 2.png

Run

python main.py

Output

A window will display a 6-panel visualization:

• Original RGB frame

• 3D K-means cluster segmentation

• ORB feature visualization

• First-pass inlier/outlier matches

• Second-pass refined inliers

• Final detected dynamic regions

Notes

• Script assumes a static camera with predefined intrinsics

• Adjust intrinsics in the script if needed

• You can change test_image_set_idx to use other folders (e.g. test3/)

Dataset

This project uses data from the TUM RGB-D dataset, specifically the freiburg3_walking_rpy sequence.

J. Sturm, N. Engelhard, F. Endres, W. Burgard and D. Cremers,
"A benchmark for the evaluation of RGB-D SLAM systems",
Proc. of the International Conference on Intelligent Robot Systems (IROS), 2012.