A real-time, Python-based toolkit for intelligent vehicle analytics:
- Vehicle detection
- Brand & model classification
- Color recognition
- Persian (Iranian) license-plate OCR + city lookup
- Accident detection with audio alerts
Built on YOLO11 & YOLO8 & ResNet18, OpenCV, and PyTorch. Ideal for smart surveillance, traffic analysis, parking management, and smart-city applications.
- Vehicle Detection
• YOLO11 car detector - Brand & Model Classification
• ResNet18 classifier trained on 27 Iranian makes/models - Color Recognition
• ResNet18 classifier for 12 common car colors
• Color-coded bounding boxes & labels - License-Plate Recognition (ANPR)
• YOLO11 plate detector + character detector
• Parses 8-character Persian/Iranian plates
• Looks up city/region from plate data - Accident Detection & Alert
• YOLO11 accident detector
• Red bounding box + “Accident (0.XX)” label
• Plays configurable beep (beep.mp3) at alert intervals - Real-Time Image & Video Support
• Live FPS counter
• Saves annotated images & video clips with timestamped filenames - Modular & Configurable
• CentralizedPATHS,THRESHOLDS,IMG_SIZE, etc.
• Easily swap models or adjust thresholds - Classes of cars
"Arisan", "Atlas", "Dena", "L90", "Mazda vanet", "Megan", "Neissan", "Pars", "206", "206 SD", "207", "405", "Peykan", "Pride", "Pride vanet", "Pride 111", "Quik", "Rana", "Rio", "Saina", "Samand", "Shahin", "Soren", "Tara", "Tiba", "Tiba 2", "Zantia"
- Python 3.8+
- (Optional) CUDA-enabled GPU for real-time performance
pip install torch torchvision ultralytics opencv-python pandas playsound pillow numpyIt’s recommended to use a virtual environment.
weights/ directory has:
weights/
├── car_det_model.pt
├── plate_det_model.pt
├── car_name_model.pth
├── color_model.pt
├── char_model.pt
└── accident_model.pt
precision recall f1-score support
Arisan 0.884 0.910 0.897 267
Atlas 0.960 0.930 0.944 128
Dena 0.990 0.963 0.976 1007
L90 0.439 0.795 0.566 73
Mazda-2000 0.997 0.995 0.996 875
Megan 0.874 0.854 0.864 130
Neissan 0.995 0.994 0.995 878
Peugeot Pars 0.997 0.977 0.987 1685
Peugeot-206 0.987 0.940 0.963 1658
Peugeot-206 SD 0.614 0.780 0.687 173
Peugeot-207 0.999 0.988 0.994 1632
Peugeot-405 0.991 0.972 0.981 1299
Peykan 0.995 0.993 0.994 1063
Pride 0.991 0.979 0.985 1402
Pride 111 0.992 0.977 0.984 1216
Pride vanet 0.764 0.915 0.833 117
Quik 0.994 0.982 0.988 1429
Rana 0.949 0.943 0.946 335
Rio 0.985 0.995 0.990 200
Saina 0.877 0.916 0.896 179
Samand 0.996 0.990 0.993 1735
Shahin 0.655 0.907 0.761 86
Soren 0.711 0.869 0.782 99
Tara 0.929 0.929 0.929 183
Tiba 0.895 0.873 0.884 157
Tiba 2 0.985 0.981 0.983 1193
Zantia 0.434 0.930 0.592 57
accuracy 0.970 19256
macro avg 0.884 0.936 0.903 19256
weighted avg 0.976 0.970 0.972 19256
Accuracy: 0.9699314499376818
Precision (macro): 0.8844069607888646
Recall (macro): 0.9361670722807696
F1 (macro): 0.9033185541147517
Plates/city_plateinfo.txt (UTF-8 CSV) format:
letter,number,city
a,11,Tehran
b,12,Isfahan
...
Place your beep file at sound/beep.mp3.
All major settings live in img & video detection.py:
PATHS = {
"car_det_model": "weights/car_det_model.pt",
"plate_det_model": "weights/plate_det_model.pt",
"car_name_model": "weights/car_name_model.pth",
"color_model": "weights/color_model.pt",
"char_model": "weights/char_model.pt",
"accident_model": "weights/accident_model.pt",
"city_plateinfo": "Plates/city_plateinfo.txt",
"alert_sound": "sound/beep.mp3",
"source": "INPUT.JPG", # or "input.mp4" / camera index
"output": "output",
}
IMG_SIZE = (220, 165) # ResNet image size
THRESHOLDS = {
"car_name": 0.65,
"car_color": 0.70,
"accident": 0.50
}
FONT_SCALE = 0.5
THICKNESS = 1
ALERT_INTERVAL_SEC = 2 # Min seconds between beepspython img & video detection.py
# Make sure PATHS["source"] points to .jpg/.png- Displays annotated image
- Saves to
output/output_YYYYMMDDhhmmss.jpg
python img & video detection.py
# Set PATHS["source"] = "input.mp4" or camera index (0, 1, …)- Opens a real-time window with FPS
- Saves annotated video to
output/output_YYYYMMDDhhmmss.mp4 - Press q to quit
You can run live vehicle detection, recognition, and analytics using your computer’s webcam. The system will process each video frame in real-time, annotate detections and predictions, and (optionally) record > the result as a video file.
1-Set the Source
In Live.py , set the video source to your webcam.
cap = cv2.VideoCapture(0)
line 261, 0 is default webcam
2- Run the Script
For the default webcam:
python Live.py
.
├── Live.py
├── img & video detection.py
├── Fine-tune/
│ └── car name model.ipynb
├── weights/ # Model checkpoint files (6 files)
├── Plates/
│ └── city_plateinfo.txt
├── sound/
│ └── beep.mp3
├── INPUT.JPG / input.mp4 # Example inputs
├── output/ # Auto-created outputs
└── README.md
Below are example outputs from the system.
click >> video output
- Vehicle detection in video
- Vehicle brand and color recognition
- License plate recognition
 |
 |
|---|---|
| Vehicle detection in the presence of occlusion | Overlapping vehicles detection |
 |
 |
|---|---|
| Accurate license plate recognition | License plate city identification (Iranian plate) |
 |
|---|
| Another example of license plate reading |
 |
|---|
| Accident detection in real-time traffic footage |
Our models were trained on a combination of:
- Custom Scraped Images from Iranian automotive platforms:
- Iran Khodro (ایران خودرو)
- Saipa (سایپا)
- Divar (دیوار)
- Bama (باما)
- Other online classifieds & dealer sites
- Public Roboflow Collections
- Car detection (various makes/models)
- License-plate detection & OCR
- Accident/incident examples
- Annotation & Augmentation
- YOLO-formatted bounding boxes & labels
- Standardized color, model & plate classes
- Augmentations: scaling, rotation, lighting, blur
- Fork the repository
- Create a feature branch (
git checkout -b feature/xyz) - Commit your changes (
git commit -m "Add new feature") - Push to your fork (
git push origin feature/xyz) - Open a Pull Request
Distributed under the MIT License. See LICENSE for details.
vehicle-detection · car-recognition · car-color-detection · license-plate-recognition · persian-plates · anpr · accident-detection · yolo11 · pytorch · opencv · deep-learning · smart-city · traffic-analysis · multitask-learning
Github.com/RezaGooner