Disability Prediction System using YOLO

Overview

This project is an end-to-end computer vision application for disability-related object detection using the Ultralytics YOLO framework. It includes:

• Model training, evaluation, and batch prediction scripts
• A full-featured PyQt5 desktop application for real-time webcam inference and single-image prediction
• Analytics dashboards with inference-time, confidence, and detection statistics
• CSV logging and visual output storage

The system is designed for research, experimentation, and demonstration purposes, with optional GPU acceleration.

---

Project Structure

project-root/
│
├── application.py        # PyQt5 GUI application (image + webcam inference)
├── ss.py                 # Utility script to locate Python Scripts folder
├── TRAIN.pdf             # Contains TRAIN.py, EVALUATE.py, PREDICT.py source
├── predictions/          # Saved prediction images (auto-created)
├── prediction_log.csv    # Inference log (auto-generated)
├── app_errors.log        # Application error logs
└── results.zip           # Training or evaluation artifacts (optional)

---

Key Features

Desktop Application (application.py)

• Load YOLO .pt models dynamically

• Real-time webcam detection with frame skipping for low latency

• Single-image prediction support

• Live analytics dashboard:

  • Inference time per prediction
  • Average confidence score
  • Object count per frame

• CSV logging of all predictions

• GPU acceleration (CUDA) when available

Model Lifecycle Scripts (from TRAIN.pdf)

• Training with configurable hyperparameters
• Evaluation with mAP, precision, and recall metrics
• Batch Prediction on image directories
• ONNX export for deployment

---

Requirements

Software

• Python 3.8+
• Windows (recommended for PyQt5 webcam backend)

Python Dependencies

ultralytics
opencv-python
torch
torchvision
numpy
pandas
matplotlib
PyQt5

Install dependencies:

pip install -r requirements.txt

(Create requirements.txt from the list above if not already present.)

---

Running the GUI Application

python application.py

Application Workflow

1. Launch the application

2. Click Load Model and select a YOLO .pt file

3. Choose one of the following:

   • Predict on Image (static image inference)
   • Start Video (real-time webcam detection)

4. View analytics and saved outputs in real time

All predictions are logged to prediction_log.csv.

---

Training the Model

Create a training script based on TRAIN.py:

python train.py

Training configuration:

• Dataset: disability.yaml
• Epochs: 50
• Optimizer: AdamW
• Image size: 640
• Batch size: 16

Model artifacts are saved under:

runs/detect/train*/

---

Evaluating the Model

python evaluate.py

Outputs include:

• mAP@50
• mAP@50–95
• Precision
• Recall
• Per-class AP metrics

---

Batch Prediction

python predict.py

• Input directory: test/images/
• Output directory: predictions/

Each image is processed and saved with bounding boxes overlaid.

---

Hardware Acceleration

• Automatically detects CUDA-enabled GPUs
• Falls back to CPU if GPU is unavailable
• Half-precision inference (FP16) enabled when supported

---

Logging and Outputs

• prediction_log.csv – structured inference metrics
• predictions/ – annotated images
• app_errors.log – runtime error tracking

---

Notes

• Webcam performance depends on camera quality and system resources
• Frame skipping is enabled to reduce latency
• ONNX models are supported for inference and evaluation

---