Child Detection Model Using YOLOv5

Abstract

Closed-circuit television (CCTV) surveillance systems have been installed in public locations to search for missingchildren and fight crime. The Penang City Council has deployed a face recognition CCTV monitoring system. As a result, this research aims to identify children who were in the wrong area or at the wrong time and then notify authorities such as police and parents. According to child detection research, the average child loss rate is more significant due to lacking child detection features. Existing research employs machine learning and deep learning across several platforms, yielding inaccurate accuracy findings. Using the YOLOv5 algorithm, this study will categorize images based on children's detection in restricted locations. Coco, Coco128, and Pascal VOC were chosen because they are the standard datasets of YOLOv5 and the public dataset INRIA Person. Annotations and augmentation techniques are employed in the pre-processing phase to acquire labeling in text file format and offer data for any object position. The YOLOv5s model will then be designed to make the proposed detector model. After training using YOLOv5s, a child detector model is produced and evaluated on the dataset to acquire findings according to recall, precision, and mean average precision (mAP) performance metrics. Finally, the performance metrics obtained from all four datasets are compared. The INRIA Person dataset performed the best, with a recall of 0.995, an accuracy of 0.998, and a mean Average Accuracy of 0.995. Nevertheless, the findings for both YOLOv5s and the proposed model are pretty close. This demonstrates that the proposed model can detect as well as the YOLOv5s model.