live people counting

Final polish: accurate-only pipeline + YOLOv11 custom backend support + clean compact UI.

What was improved in this update:

• switched to accurate-only detection profile (removed Fast/Balanced selection)
• added optional integration path for B4rtekk1 YOLOv11 backend when local repo+weights are present
• kept safe automatic fallback to Ultralytics YOLO11 weights if custom backend is unavailable
• boosted runtime smoothness with adaptive detection interval and OpenCV optimizations
• fixed IN/OUT ambiguity by adding explicit IN-side selection during setup
• improved line crossing logic to use the selected IN side instead of guessing
• cleaned HUD back to a compact v07-style layout (small text, simple top-left panel)

Also removed temporary testing artifacts from final output:

• removed local video test file and MJPEG helper server script
• removed extra dependency used only for YouTube/file stream testing

Current final build is stable for webcam/RTSP inputs with manual door-line calibration and reliable IN/OUT counting.
Setup no longer asks for performance profile; accurate mode is always used.

Update: local video playback is now a core camera source for the final project.

What changed in this follow-up:

• promoted Local Video File from test wording to full source option in setup
• aligned setup/documentation so webcam, RTSP, and local video are all first-class sources
• local video can now be used for regular counting workflows, demos, and offline validation

cleaned junk code.

Pre-release cleanup and merge stage.

Pulled in the polished app structure:

• cleaner config + tracker behavior
• more complete overlays and runtime logging
• setup flow and counting logic aligned with final codebase

This stage is basically the last checkpoint before packaging docs/license and calling it done.

improved code to make it smooth.

Performance-focused pass.

What I changed:

• added frame skipping in config and main loop
• reused last detection result between skipped frames
• forced low buffer size and target FPS settings on camera

Tradeoff is obvious: a bit less accurate when people move fast, but much smoother on weak hardware. For this project, stable realtime view is more important than perfect per-frame detection.

improved active track count and event log print format.

Worked on counting quality and visibility.

Changes in this stage:

• added active track count to the overlay
• improved event log print format with timestamps
• kept tracker API a bit cleaner (get_active_persons, reset_counts)

Annoying part: when a few people cross together, nearest-neighbor still swaps IDs sometimes. Not fully solved yet, but output is easier to debug now because logs are clearer.

camera selecting ability and door selecting ability.

Usability pass: removed hardcoded assumptions.

Implemented setup flow:

• choose webcam or RTSP URL
• test camera quickly
• draw door line with mouse drag

Mouse callback logic was annoying at first (line would jump because drag state was wrong). Fixed by handling LBUTTONDOWN/MOVE/UP carefully.

This stage feels like a real app now, not just a dev prototype.

In / out printed with fixed door line at mid-frame.

Built a basic nearest-neighbor tracker and connected it to a fixed door line at mid-frame.

Now it can print IN / OUT events and keep counters on screen.

Big issue today: jitter around the line created fake crossings. I didn’t fully solve it here, but at least track aging/removal is in place so stale IDs don’t pile up forever.

person detection.

Added YOLO person detection. The first run looked like a hang, but it was just downloading model weights.

Current behavior:

• reads webcam frame
• runs detector
• draws person boxes with confidence

Still rough. Box flicker is obvious and FPS drops on slower machines, but this is enough to visually verify detections before I touch tracking.

webcam opening.

Started from scratch and forced myself to keep it tiny: just open webcam, show frames, exit on ESC.

I wasted a bit of time because camera index 0 didn’t work on first try (USB cam was on 1). I added the index prompt so I don’t keep editing code every run.

No ML yet, no counting yet. Just confirmed OpenCV loop is stable and cleanup (release + destroyAllWindows) is working.