🚑 Fall Detection System

This project implements a computer vision-based fall detection system using deep learning techniques. The system can detect whether a person has fallen in images or video frames.

📁 Project Structure

• main.py - Main script for running fall detection and evaluation
• data_handler.py - Handles dataset loading and preprocessing
• my_models.py - Contains model definitions
• fall_dataset/ - Directory containing training and validation data
• images/ - Contains input images
• labels/ - Contains corresponding labels
• model_weights/ - Directory containing trained model weight files (.h5 files)

📊 Data Source

The dataset used in this project is sourced from Kaggle.

📄 Dataset Details

• Name: Fall Detection Dataset
• Link: https://www.kaggle.com/datasets/uttejkumarkandagatla/fall-detection-dataset
• Size: 374 samples
• Description: Images of people in various postures including standing, sitting, and fallen states in different environments

🧠 Models

The system uses two main models:

1. Fall Detection Model (fall_detection_model) - A custom deep learning model that classifies cropped person images as fallen or not fallen.
2. YOLO Model (yolo_model) - Uses YOLOv5 to detect people in images and generate bounding boxes.

📦 Dataset Labels

The Fall_dataset class processes the dataset with the following label convention:

• 0 - Fallen
• 1 - Standing
• 2 - Sitting

🚀 Usage

🔍 Running Evaluation

This will:

• Load the fall detection model with pre-trained weights
• Process validation images with YOLO to detect people
• Apply the fall detection model to each detected person
• Calculate performance metrics (accuracy, precision, recall, F1 score)

🎓 Training a New Model

The system supports training using the train method in the fall_detection_model class.

📈 Performance Metrics

The system calculates:

• Accuracy: (TP + TN) / (TP + FP + TN + FN)
• Precision: TP / (TP + FP)
• Recall: TP / (TP + FN)
• F1 Score: 2 * TP / (2 * TP + FP + FN)

Where:

• TP = True Positives (Fall - Detected)
• FP = False Positives (Not Fall - Detected)
• TN = True Negatives (Not Fall - Not Detected)
• FN = False Negatives (Fall - Not Detected)

🔧 Requirements

• TensorFlow
• PyTorch
• OpenCV
• Matplotlib
• NumPy
• YOLOv5 (via torch hub)
• PIL