feature_extractor

Object tracking

Object tracking using OpenCV features detectors (detectors) and descriptor extractors (descriptors) algorithms with GUI for fun, tests and education.

This application only processes one image. Good object tracker processes sequence of images in a video. For example, this OpenCV object tracking algorithms.

Only OpenCV integrated detectors, descriptors and detector-descriptors are used.

Neural Network detector-descriptors (such as R2D2, D2NET, SUPERPOINT, ORB-SLAM2, DELF, CONTEXTDESC, LFNET, KEYNET, DISK, etc) and descriptors (such as TFEAT, HARDNET, GEODESC, SOSNET, L2NET, LOGPOLAR, etc) are not considered.

All object trackers in the application are placed in decrease of efficiency and implemented rotation invariant and scalable except of "StarDetector + DAISY". It doesn't mean that lower methods are always ineffective! However, for this task it is so, because there is no "silver bullet" method for all tasks.

All feature detector-descriptor logic is in the logic_extractor.py file. Snapshots, logs and configuration parameters are saved in temp directory of the current feature_extractor folder. Overall the GUI source code is not as elegant as I would like, but it works :-).

Previous simple script is here SIFT object tracking. SIFT algorithm became free since March 2020. SURF algorithm is patented and is excluded from OpenCV. Now SURF is for Python version 3.4.2.16 and older.

App is tested on Windows 10 for Python 3.12.

External libraries:

• OpenCV to process images.
• NumPy to support calculations with arrays.
• Pillow to open images of various formats.

How to start the app:

# Install additional libraries
pip install -r requirements.txt
# Run the application
python runme.py

Usage:

1. To open the GUI, run command in you console python runme.py.
2. Place object in front of the web camera, so it take all visible space.
3. Press Get snapshot button. Application will make snapshot of the object to track.
4. After taking snapshot there will be a red rectangle around tracking object and green lines connecting special keypoints of the image.

Note: Rectangular object, like book, is tracked better than face.

Check out the PowerPoint presentation of the application in the data subdirectory.