Beyond faces: Clothing attribute‐based missing child retrieval in video surveillance

Abstract

Clothing attributes serve as distinctive and immediately recognizable features, which are utilized to identify individuals in video surveillance footage. A novel framework, the ByteTrack Child Detection Network (BT‐CDNet), integrated with a Clothing Attribute Recognition (CAR) model is proposed to detect and retrieve missing children based on clothing attributes. The framework employs YOLOv8s with the Convolutional Block Attention Module (CBAM) and the ByteTrack algorithm for detection and tracking, while the CAR model performs multi‐attribute recognition to predict clothing attributes and identify the best match. The Children Clothing Attributes (C2A) dataset is introduced to improve clothing identification in surveillance images and capture the unique clothing characteristics of children. CAR significantly outperforms existing methods, achieving a mean Average Precision (mAP) of 93.2% and Rank‐1 accuracy of 95.6% on the benchmark dataset SoftBioSearch and a mAP of 91.51% and an inference speed of 48.6 frames per second on the C2A dataset.