An Enhanced Framework for Monitoring Toddlers’ Safety in Smart Homes Using Deep Learning

Abstract

Monitoring toddlers in smart homes is essential for ensuring their safety due to their unpredictable behavior and the varying nature of home environments. This dissertation presents a novel toddler detection and tracking system that surpasses existing methods by employing an optimized Convolutional Neural Network (CNN). The developed system is capable of detecting and tracking multiple toddlers simultaneously, offering a scalable architecture that can easily integrate future Mask R-CNN advancements. Implemented using Python, TensorFlow, and Keras, and structured with the Object-Oriented Analysis and Design (OOAD) methodology, the system achieves a remarkable accuracy of 98.94%. This significant improvement demonstrates its superior robustness and effectiveness in handling dynamic and diverse home environments. The enhancements made include the fine-tuning of CNN parameters to better accommodate various lighting conditions and occlusions commonly found in home settings. Real-time preprocessing ensures that the system maintains high accuracy without compromising processing speed, making it suitable for continuous monitoring applications. Transfer learning with MobileNetV2 allows the system to leverage pre-trained models, accelerating the training process and improving generalization across different home layouts. The system's high accuracy of 98.94% not only highlights its effectiveness but also positions it as a reliable solution for enhancing toddler safety in smart homes.