AI Pipeline for Real-Time Child Labour Detection and Age Estimation
A team of researchers has created a real-time computer-vision system aimed at improving Child Labour Monitoring and Remediation Systems (CLMRS) by delivering ongoing, presence-based data. This prototype tackles the issue of under-detection affecting approximately 138 million children worldwide. Utilizing the CerberusDet framework, the system features a multi-task person and face detector based on YOLO26x, age estimation through a combination of MiVOLO v2 and a child-specialist model for ages 0-12, alongside ByteTrack tracking, ArcFace, and DINOv2 for re-identification. The detector enhances person mAP@0.5 from 0.390 to 0.683 compared to the previous baseline, while the child specialist records a mean absolute error of 1.944 years in validation. Published on arXiv with identifier 2608.14770, the study emphasizes AI's potential for more precise monitoring, although ethical and practical challenges persist. The system is intended only as a research prototype, with no plans for immediate implementation.
Key facts
- An estimated 138 million children are in child labour worldwide.
- The pipeline combines YOLO26x backbone in CerberusDet framework for detection.
- Age estimation uses MiVOLO v2 and a child-specialist model for ages 0-12.
- ByteTrack tracking and ArcFace/DINOv2 re-identification are integrated.
- Detector improves person mAP@0.5 from 0.390 to 0.683.
- Child specialist achieves 1.944 years MAE on children-only validation.
- The system is a research prototype, not deployed.
- Published on arXiv with ID 2608.14770.
Entities
Institutions
- arXiv