Real-Time Video Analytics Pipeline

A production-style computer vision pipeline for live video streams — webcam, IP camera, or RTSP feed. Built with YOLOv8 and OpenCV.

---

Features

• Real-time object detection — YOLOv8 inference on every frame with bounding boxes, class labels, and confidence scores
• Alert zone — define a polygon region; triggers a visual alert when a target class enters it
• Stream reconnection — automatically recovers from dropped RTSP/IP camera connections
• FPS counter — live performance monitoring overlaid on the video
• Per-class object counts — real-time HUD showing how many of each class are detected
• Flexible source — works with webcam, video file, RTSP stream, or YouTube stream URL

---

Demo

[FPS  28.4]  objects: person=3, car=1  |  zone: ALERT
[FPS  29.1]  objects: person=2         |  zone: clear

---

Project Structure

rtsp_pipeline/
├── pipeline.py       # main loop: capture → inference → display
├── detector.py       # YOLOv8 wrapper + zone intrusion logic
├── config.py         # all settings in one place
└── requirements.txt

---

Quickstart

1. Clone the repo

git clone https://github.com/YOUR_USERNAME/RTSP-IP-Camera-Real-time-Object-Detection.git
cd RTSP-IP-Camera-Real-time-Object-Detection

2. Install dependencies

pip install -r requirements.txt

3. Run

# Default: opens webcam
python pipeline.py

# Custom source via argument
python pipeline.py --source "rtsp://admin:pass@192.168.1.64:554/stream"
python pipeline.py --source "video.mp4"
python pipeline.py --source "https://yt-dlp-stream-url"

Press Q or Escape to quit.

---

Configuration

All settings are in config.py — no need to touch the source code.

Parameter	Default	Description
SOURCE	0	Webcam index, RTSP URL, or video file path
MODEL	yolov8n.pt	YOLOv8 variant — n (fastest) → x (most accurate)
CONFIDENCE	0.40	Minimum detection confidence (0.0 – 1.0)
IOU_THRESHOLD	0.45	NMS overlap threshold for duplicate box removal
IMGSZ	640	Inference resolution
ALERT_ZONE	centre rect	Polygon as fractions of frame size (0.0 – 1.0)
ALERT_CLASSES	["person"]	Classes that trigger zone alert. Empty = all classes

Example: RTSP IP camera

SOURCE = "rtsp://admin:password@192.168.1.64:554/Streaming/Channels/101"

Example: custom alert zone (bottom-left corner)

ALERT_ZONE = [
    (0.0, 0.7),
    (0.4, 0.7),
    (0.4, 1.0),
    (0.0, 1.0),
]

Example: alert on any class

ALERT_CLASSES = []   # empty = all classes trigger the alert

---

Using a YouTube stream (for testing)

pip install yt-dlp

# Get stream URL
yt-dlp -f best -g "https://www.youtube.com/watch?v=VIDEO_ID"

# Pass it directly
python pipeline.py --source "https://..."

Note: YouTube stream URLs expire after a few hours. Re-run yt-dlp to get a fresh one.

---

Model Size vs Speed Tradeoff

Model	Size	Speed (CPU)	Use case
yolov8n.pt	6 MB	~30 FPS	Real-time on CPU
yolov8s.pt	22 MB	~20 FPS	Balanced
yolov8m.pt	52 MB	~10 FPS	Higher accuracy, GPU recommended
yolov8l.pt	87 MB	~5 FPS	GPU only

Models download automatically on first run.

---

Key Engineering Decisions

Why fractional zone coordinates? Zone stays correct regardless of resolution — same config works at 480p and 4K without any changes.

Why reconnection logic? Real RTSP streams drop packets and lose connection on network hiccups. Production systems must recover automatically without human intervention.

Why FPS averaged over 1 second? Per-frame FPS fluctuates wildly (28ms one frame, 35ms the next). Averaging over a full second gives a stable, readable number.

Why defaultdict(int) for counts? Avoids manual key existence checks — automatically initializes any new class to 0 before incrementing.

---

Requirements

• Python 3.8+
• OpenCV 4.8+
• Ultralytics (YOLOv8)
• NumPy

---

License

MIT