streamware

YOLO11n + ByteTrack Tracking Benchmark

Standalone demo for testing the recommended AMD CPU/iGPU tracking stack:

• YOLO11n with OpenVINO acceleration (or PyTorch)
• ByteTrack via Supervision library
• MOG2 motion gating to reduce detection frequency
• Track state management (new/stable/lost)
• Threaded RTSP capture for live streams

Benchmark Results (AMD Ryzen 7940HS)

Tested on 640x480 RTSP stream @ 10 FPS:

Metric	Value
YOLO11n detection	~10-12ms (PyTorch)
Full pipeline	74+ FPS capability
Motion gate savings	45-86% fewer detections
Track events	Enter/leave detection working

Key findings:

• PyTorch is already very fast (~10ms), OpenVINO optional
• Motion gating is highly effective
• ByteTrack provides stable track IDs
• FPS limited by source stream, not detection

Quick Start

# Install dependencies
pip install -r requirements.txt

# Test with webcam (with display + stats overlay)
python tracker_demo.py --source 0 --display --stats --trace

# Test with RTSP stream (15 second benchmark)
python tracker_demo.py \
    --source "rtsp://user:pass@ip:554/stream" \
    --no-openvino --duration 15 --stats

# Process video file and save output
python tracker_demo.py \
    --source input.mp4 \
    --output result.mp4 \
    --stats

Key Options

Option	Description	Default
--source, -s	Video source (file, RTSP URL, webcam index)	required
--display, -d	Show live window	off
--output, -o	Save output video	none
--model, -m	YOLO model path	yolo11n.pt
--no-openvino	Disable OpenVINO (use PyTorch)	off
--imgsz	Detection input size	640
--conf	Confidence threshold	0.5
--motion-threshold	Pixels to trigger detection	1000
--periodic-interval	Force detection every N frames	30
--track-buffer	Frames before deleting lost track	90
--trace	Show movement traces	off
--stats	Show FPS/detection stats overlay	off
--duration	Limit benchmark to N seconds	0 (unlimited)
--max-frames	Limit benchmark to N frames	0 (unlimited)

Keyboard Shortcuts (with --display)

• q - Quit
• s - Save screenshot

Tuning Tips

Too many false detections?

• Increase --conf to 0.6-0.7
• Increase --motion-threshold to 2000-3000

Tracks lost too quickly?

• Increase --track-buffer to 120-150 (4-5 sec at 30 FPS)
• Decrease --periodic-interval to 15-20

Low FPS?

• Use --imgsz 480 for faster detection
• Increase --motion-threshold to reduce detection frequency

Integration Recommendations

Based on benchmark results, for stream-ware integration:

1. Replace ObjectTracker with ByteTrack via Supervision
2. Keep PyTorch backend (fast enough, simpler than OpenVINO)
3. Wire motion gating from FrameDiffAnalyzer to control YOLO calls
4. Use track state events (new/lost) for TTS announcements

This site is open source. Improve this page.