This report documents the development of a real-time object detection module for early childhood learning, designed to address the lack of toddler-centric datasets and mobile-optimized, privacy-preserving detection systems. Following the Structured Engineering Thesis Framework (SETF), the project integrates problem definition, dataset engineering, model training, and evaluation into a coherent research lifecycle. The system is built on lightweight YOLOv11n and YOLOv11s models trained on a combined dataset comprising a custom toddler collection, COCO, and multiple fruit and vegetable datasets. The training pipeline incorporates label harmonization, annotation normalization, targeted augmentation, and class imbalance mitigation across more than 150 child-relevant categories. Preliminary benchmark-oriented trials indicate that YOLOv11s achieves the strongest accuracy–latency balance with an estimated mAP@50 of 0.86, while YOLOv11n offers faster inference for lower-end devices. The module is designed for on-device inference, ensuring offline operation, low latency, and privacy preservation. This work demonstrates that lightweight edge-deployed vision models can effectively support early learning by enabling children to interact with real-world objects through immediate visual and auditory feedback. This work was conducted at Arab International University (AIU), Syria. The official website of the university is: https://www.aiu.edu.sy
Aya Abdeen (Sun,) studied this question.