Towards Enhancing Child Safety: A Deep Learning Approach to Detect Child Safe and Unsafe Objects

The safety and well-being of children are a paramount concern, and the accurate detection of potentially haz-ardous objects is crucial in ensuring their protection, especially in indoor situations. In the Artificial Intelligence (AI) age, the future belongs to domestic robots; they will be used for child education, surveillance, and healthcare support. Early identification of unsafe objects and intervention by these next-generation social family robots can play an essential role in providing a safer environment for children. Keeping these issues in mind, this study aims to tackle daily risks and accidents that endanger children by using computer vision to detect unsafe objects. To achieve this goal, a unique dataset namely ChildSUn is introduced containing approximately 5350 images of child-safe and unsafe objects of ten categories. Various CNN architectures, including custom CNN, VGG16, VGG19, ResNet50, Inception-V3, EfficientNet-B0, and MobileNetV2, are implemented on the prepared dataset. A deep dive analysis is performed on the implemented models. The EfficientNet-B0 model showed exceptional performance, reaching a remarkable level of accuracy of 99.63% on the test dataset. These findings underscore the potential of EfficientNet-B0 as an effective solution for real-world child safety applications and encourage further exploration of deep learning techniques to enhance child protection globally. We hope this work will pave the way for future social household robot development, which will ensure child safety in any place.