I am building a real-time vehicle recognition system using Python.
The system should support these input methods:

1. Local video file upload

2. Live camera feed using a webcam

3. camera URL

4. YouTube video URL

Additionally, I want to implement:

• OpenCV point/ROI selection (mouse click points on frame)

• Zone detection (detect vehicles only inside selected polygon/zones)

• A modular architecture so each input method follows the same processing pipeline.

I am able to load videos individually, but I am struggling to design a clean structure that handles all input types consistently. I also need guidance on how to integrate OpenCV cv2.setMouseCallback() for ROI selection before starting the detection loop.

Questions:

1. What is the best way to structure a common video input pipeline for:

   • file path

   • webcam

   • RTSP URL

   • YouTube URL (via pytube or yt-dlp)?

2. How do I properly implement OpenCV point selection (mouse clicks) on the first frame, and pass those zone coordinates to my detection loop?

3. Are there standard patterns for combining ROI selection + detection + frame-by-frame processing in real-time systems?

4. Any example architecture or code pattern that cleanly separates:

   • input handler

   • processing/detection (YOLO/OpenCV)

   • output/logging