Artificial Intelligence as a Tool for Combating Child Labour: A Real-Time Edge Vision Pipeline for Child Detection and Age Estimation

Mark Nowak

Source-linked AI summary

Artificial Intelligence as a Tool for Combating Child Labour: A Real-Time Edge Vision Pipeline for Child Detection and Age Estimation

Mark Nowak

arXiv:2608.14770v1Aug 14, 2026cs.CVcs.AI

TL;DR

Periodic household visits and interviews under-detect child labour, motivating continuous workplace evidence for CLMRS. This research prototype combines real-time detection, child-focused age estimation, tracking, re-identification, and fusion to produce reviewable per-person records. In a Zimbabwe pilot, software tuning improved yield 36-fold, while consolidation reduced over-reporting from 9.1x to 1.8–3.9x with zero proven-false merges.

Problem
The paper asks whether computer vision can provide CLMRS with continuous, presence-based workplace observations combining child-grade age estimation, identity persistence, and real-time operation.
Method
The paper builds a real-time pipeline combining person and face detection, cascaded age estimation, tracking, re-identification, and track-level fusion for reviewable records.
Results
36-fold software-tuning yield improvement reduced over-reporting from 9.1 times the attendance register to 1.8–3.9 times, with zero proven-false merges.
Takeaways & Limitations
Camera geometry and labelled field validation are identified as priorities for future deployment and for separating discovery from estimator bias.
Takeaways & Limitations
Field age evidence is unverifiable because small, backlit profile faces make neighbouring age bands visually indistinguishable to humans.

Abstract

from arXiv · show

An estimated 138 million children remain in child labour worldwide, and the monitoring systems used by affected sectors, built on periodic household visits and interviews, systematically under-detect them. We present a real-time computer-vision pipeline, built and operated solely as a research prototype, that studies the feasibility of giving Child Labour Monitoring and Remediation Systems (CLMRS) a continuous, presence-based evidence channel. The pipeline combines a multi-task person and face detector (YOLO26x backbone in the CerberusDet framework), cascaded age estimation pairing MiVOLO v2 with a child-specialist model for ages 0-12, ByteTrack tracking, ArcFace and DINOv2 re-identification, and track-level fusion producing reviewable per-person records. The detector raises person mAP@0.5 from 0.390 to 0.683 over the previous-generation baseline; the child specialist reaches 1.944 years MAE on children-only validation, where widely used open-source stacks err by 18-23 years. FP8 TensorRT compilation yields a 1.77x speedup at +0.002 years MAE, bringing the pipeline above twice real-time on embedded hardware. On 26.8 hours of proxy video the system finds 634 unique child candidates versus 285 for its predecessor. We further report a seventeen-day unattended field pilot on a farm in Zimbabwe (38.7 million frames, six cameras) evaluated against a daily attendance register: software tuning improved detection yield 36-fold, and identity consolidation under a simultaneity veto cut over-reporting from 9.1x to 1.8-3.9x with zero proven-false merges. We document training and quantisation failures alongside successes, and the data-protection and human-in-the-loop safeguards such a system requires.

1 Introduction

The paper presents a research-only computer-vision prototype intended to add continuous, presence-based workplace observations to CLMRS, addressing persistent gaps in child-labour monitoring. It combines detection, child-focused age estimation, tracking, re-identification, edge deployment, and multi-level evaluation, while documenting limitations and safeguards.

Motivation: 138 million children aged 5–17 remain in child labour worldwide, including 54 million in work directly endangering health or safety.The long-term trend has improved, but the Sustainable Development Goal 8.7 target of full elimination by 2025 was missed.
Monitoring gap: CLMRS reached 1.17 million cocoa-farming households by September 2024, covering about 55 per cent of cocoa-growing households in Côte d’Ivoire and Ghana.The paper frames computer vision as a continuous workplace signal feeding candidate sightings into existing CLMRS verification and remediation workflows.
System design: The pipeline combines YOLO26x-CerberusDet person-and-face detection, cascaded MiVOLO v2 age estimation with a 0–12 child specialist, ByteTrack, ArcFace, DINOv2, and CLMRS-compatible reporting.Its perceptual tasks are locating people and faces, estimating age, and maintaining identities across frames and camera views.
Model development: 1.944 years MAE was achieved by the child specialist on children-only validation for ages 0–12, despite catastrophic forgetting, distribution mismatch, and oversampling failures.The introduction explicitly reports negative training results alongside the successful specialization of MiVOLO v2.
Edge deployment: 1.77× speedup was obtained from FP8 TensorRT compilation of the age cascade at a 0.002 years MAE quality cost.The paper also reports an FP8 detector engine increasing full-pipeline throughput by 18 per cent while halving engine size, with 2.1 per cent fewer detected children.
Evaluation: 122 per cent more unique child candidates were surfaced on 26.8 hours of video, while the field pilot processed 38.7 million frames from six cameras over seventeen unattended days.The pilot used a single embedded device and evaluated outputs against a daily attendance register plus a simultaneous-visibility false-merge ground truth.

2 Background and Related Work

Prior work mainly maps child-labour risk or detects proxy infrastructure at area level, while individual-level recognition systems remain relatively uncommon and narrow. The proposed background also draws on established real-time detection, age-estimation, tracking, and re-identification methods suited to surveillance conditions.

Area-Level Monitoring: Area-level tools estimate risk or locate proxy infrastructure, but they do not observe individual children.DIGICHILD uses one-square-kilometre risk grids, while satellite, drone, and aerial systems support infrastructure mapping or humanitarian monitoring.
Individual-Level Systems: Individual-level child-labour systems remain rarer and narrower, including image classifiers, CCTV child detectors, embedded adult–child classifiers, and minor-flagging systems.The cited work spans deep-learning classification, YOLOv5 detection, embedded deployment, and facial-plus-body-silhouette cues.
Real-Time Detection: YOLO-family single-stage detectors remain the default for real-time deployment, while YOLO26 targets small objects and dense scenes through architectural and training changes.The cited YOLO26 generation removes distribution focal loss, uses NMS-free inference, and adds progressive loss balancing and small-target-aware label assignment.
Age Estimation: MiVOLO jointly processes face and body crops to preserve age signals when faces are small, averted, or occluded.The passage situates this approach within the progression from CNN regression and ordinal formulations to distribution-aware, compact, and transformer models.
Tracking and Re-Identification: ByteTrack retains low-confidence boxes to preserve partially occluded people, while ArcFace embeddings and DINOv2 features support merging track fragments into unique persons.The passage contrasts ByteTrack with appearance-augmented trackers including BoT-SORT, StrongSORT, and OC-SORT, and identifies ArcFace as standard for face embeddings.

3 System Design and Methods … 3.3 Multi-Task Person and Face Detector

The system is designed as a continuously operating, edge-deployable pipeline that estimates childhood age continuously and converts frame-level detections into reviewable person records. Its multi-task YOLO26x–CerberusDet detector shares a backbone while using separate person and face heads, trained with task-aligned objectives on unified public datasets.

3.1 Design Requirements: Continuous edge operation and continuous childhood-age estimation are core requirements because target sites may lack cloud connectivity and remediation priority varies by age.The design explicitly rejects a binary adult–child decision in favor of dedicated childhood accuracy.
3.2 Pipeline Overview: The pipeline processes continuous RTSP streams in ten-minute segments using one persistent worker per camera and detects persons and faces in each frame.For persons with associated faces, the age module receives a six-channel input consisting of face and body crops.
3.2 Pipeline Overview: ByteTrack forms temporal tracks, while ArcFace and DINOv2 embeddings merge tracks belonging to the same individual within and across reappearances.Track merging uses a cosine-similarity threshold of 0.70.
3.2 Pipeline Overview: Fusion aggregates frame-level estimates into one person-level age and gender record, retaining quality flags and discarding tracks visible for less than 3.5 seconds.Tracks with fused age below the child threshold are recorded as child candidates for human triage.
3.3.1 Architecture: The detector uses a shared YOLO26x backbone within CerberusDet, branching into separate single-class, anchor-free person and face heads over P3–P5 feature levels.The backbone has 28.8M parameters, while the full model has 101.4M parameters and 366 GFLOPs at 640 × 640 input.
3.3.2 Training Objective: Training uses a composite detection loss with λbox, λcls, λdfl = (7.5, 0.5, 1.5) for both person and face tasks.The shared backbone is updated using gradients averaged across alternating task batches.
3.3.2 Training Objective: Task-aligned assignment scores candidate anchors using classification confidence and IoU, selecting the top ten anchors per ground truth as positives.Classification is trained against normalized alignment scores, linking localization quality to classification supervision.
3.3.3 Training Configuration: The detector was fine-tuned from a YOLO26x checkpoint using unified two-task public datasets for person and face detection.Person training used 108,470 training and 10,315 validation images; face training used 63,130 training and 8,174 validation images on four NVIDIA H200 GPUs.

3.4 Cascaded Age and Gender Estimation

The cascade combines cross-attended face and person streams in MiVOLO v2 with a general estimator and child specialist, routing ages conservatively to improve child-specific estimation. Training failures motivated lower learning rates and clean-label fine-tuning, yielding 1.944 years MAE for ages 0–12 on children-only validation.

Cascaded routing: The cascade uses a general model and child specialist, with gender always taken from the general model and routing thresholds of τroute = 20 years in the laboratory and τroute = 32 in the deployed pilot.The field threshold was raised because small, backlit faces made the FP8 general model collapse toward a prior of roughly 27.8 years, preventing children from reaching the specialist.
Training failures: 1 × 10−6 to 5 × 10−6 became the working learning-rate range, roughly 25 times lower than planned, after the initial 5 × 10−5 strategy degraded the model.Teacher pseudo-labels with MAE 3.65 limited the student near the teacher’s error, while self-distillation was circular.
Training redesign: The conservative re-plan improved the general model’s validation MAE from about 5.0 to 3.936 using staged warm-up and fine-tuning on clean datasets.The schedule used a frozen-backbone warm-up, then 2×10−6 on eight clean datasets and 1×10−6 on thirteen datasets.
Child-specialist results: 1.944 years MAE for ages 0–12 and 2.291 years for ages 0–18 were achieved by the child specialist on children-only validation, with CS@5 of 93.9 per cent.The specialist used 119,233 training and 19,975 validation samples, followed by a focus phase on ages 0–12.

3.5 Tracking, Re-Identification and Fusion · 3.6 Field Aggregation and Identity Consolidation

The pipeline tracks partially occluded people, fuses face- and body-based re-identification with face-backed age evidence, and classifies records using configurable child-probability thresholds. Field aggregation consolidates repeated encounters across cameras through similarity matching, union-find merging, and a simultaneity veto, with a second-pass detector recovering additional faces.

3.5 Tracking, Re-Identification and Fusion: ByteTrack retains low-confidence boxes for association, while ArcFace face and DINOv2 body embeddings merge track fragments at cosine similarity ≥0.70.Face embeddings are updated along tracks using an exponential moving average.
3.5 Tracking, Re-Identification and Fusion: Only face-backed age estimates enter the history; body-only estimates are excluded because they reproduce a dataset prior measured at 22.4 years in 89 per cent of cases.The robust age is computed as the 25th percentile of the face-backed history.
3.5 Tracking, Re-Identification and Fusion: p ≥0.5 classifies a person as a child, with laboratory thresholds τchild = 16 and amax = 19, versus field thresholds τchild = 18 and amax = 21.Field evidence packages additionally require p ≥0.55; records are tiered as high ≥0.8, medium 0.6–0.8, or low otherwise.
3.6 Field Aggregation and Identity Consolidation: Continuous multi-camera operation creates repeated daily appearances, so per-segment filing would multiply one child into many reports.The field deployment therefore adds an aggregation layer validated during the pilot.
3.6 Field Aggregation and Identity Consolidation: Daily encounters enter a shared SQLite identity registry, and matching joins encounters across cameras at face similarity 0.75 or within-window body similarity 0.72, subject to a five-year age gate.A cutoff process at local midnight assembles each day’s evidence, and the registry takes about one millisecond per segment.
3.6 Field Aggregation and Identity Consolidation: At window close, union-find merges identity groups using best-pair face similarity 0.66 or body similarity 0.61 when no usable face comparison exists.A sub-threshold face comparison vetoes the body path, and the merge is subject to a hard time guard.
3.6 Field Aggregation and Identity Consolidation: Groups cannot merge when any encounters were simultaneously visible, enforcing a group-level simultaneity veto that prevents transitive unions from creating forbidden merges.The veto enables aggressive similarity thresholds while preserving temporal consistency.
3.6 Field Aggregation and Identity Consolidation: A second-pass SCRFD face detector rescans selected person crops and recovers 2–6 per cent additional faces, while most faceless tracks remain genuinely face-invisible.The rescue pass uses reduced resolution and a bounded count.

3.7 Edge Deployment and Quantisation · 3.8 Evaluation Methodology

The study deploys quantised age-estimation services on embedded Blackwell-class hardware while documenting toolchain limitations. Evaluation spans disjoint accuracy benchmarks, an identity-free proxy-video study, and a seventeen-day Zimbabwe field pilot.

3.7 Edge Deployment and Quantisation: FP8 TensorRT deployment targets DGX Spark laboratory studies and Jetson AGX Thor field operation as containerised PyTorch, Torch-TensorRT, and TensorRT services.Both MiVOLO models use native FP8 E4M3 compilation through the Torch-TensorRT dynamo path.
3.7 Edge Deployment and Quantisation: 110 MB to 93 MB per model: FP8 quantisation shrinks each engine, while unsupported VOLO operations remain in PyTorch fallback.Maximum calibration-sample output deviation from FP32 was 5.6×10−4 on GB10 and between 10−4 and 10−3 on Thor.
3.7 Edge Deployment and Quantisation: Engines were rebuilt on Thor because TensorRT engines do not port across GPU architectures.The deployment used GB10 compute capability 12.1 and Thor compute capability 11.0.
3.7 Edge Deployment and Quantisation: FP16 compilation through the older TorchScript IR fails at runtime for VOLO, whereas the dynamo path avoids the reported shaped-tensor routing problem.FP4 quantisation could not be realised on either platform because ModelOpt fake-quantisation tensors were not traceable by torch.export.
3.8 Evaluation Methodology: Five evaluation levels use data disjoint from training wherever accuracy is scored, covering detector, age, internal cascade, film-study, and field-pilot assessments.The methodology includes benchmark comparisons, ablations, proxy-video processing, and unattended field operation.
3.8 Evaluation Methodology: 13,537 images and 156,226 ground-truth boxes form the detector benchmark comparing current and previous generations under identical GB10 FP16 settings and annotations.The combined validation set draws from COCO Person, WiderFace, CrowdHuman, and Objects365 Person.
3.8 Evaluation Methodology: 1,500 APPA-Real images and 6,165 FairFace-Africa images support age benchmarks against InsightFace and DeepFace using MAE, median absolute error, CS@5, and CS@10.Results are reported for full sets and child subsets, among other groupings.
3.8 Evaluation Methodology: 146,735 held-out samples from fourteen datasets evaluate the internal cascade and ablations by age-group MAE, CS@5, and child-model routing share.The end-to-end film study processes nineteen videos totaling 26.8 hours and 482,486 frames without ground-truth identity or age annotations.

4 Results

The results show substantial gains in detection, child-specific age estimation, candidate discovery, and unattended deployment, while quantifying throughput, gender, identity-consolidation, and camera-geometry limitations. The pilot demonstrates operational stability and improved yield, but recall remains conditional on face detection.

Detection: Person mAP@0.5 rose from 0.390 to 0.683, while detections increased 46 per cent from 1,049,992 to 1,531,954.Recall gains remove blind spots because undetected people cannot be age-estimated.
Age estimation: Child-only MAE reached 1.944 years for ages 0–12, while open-source stacks produced 18.7–22.9 years and predicted 27–35 years for children aged 5–12.The cascade improved on general-only configuration by 0.180 years MAE across children aged 0–18.
Quantisation: FP8 reduced mean latency from 22.1 to 12.5 ms per image, a 1.77× speedup, while changing MAE by +0.002 years.Per-bucket age deltas stayed below 0.006 years in either direction.
Proxy-video evaluation: 634 unique child candidates were surfaced versus 285, a 122 per cent increase, while high-confidence records grew from 49 to 69 per cent.The system reached a minimum fused age of 0.8 years and produced 213 under-10 candidates versus 23 previously.
Field pilot: 38.7 million frames were processed across six cameras without operator intervention, while software tuning raised detection yield from 0.005 to 0.181 children per segment, a 36-fold improvement.The system maintained its five-frames-per-second median target and filed every queued report successfully.
Field pilot limitations: Identity consolidation reduced register over-reporting from up to 9.1× to 1.8–3.9×, while conditional recall was 48 per cent across 595 child events.Recall excludes children whose faces were never detected, and camera geometry was identified as the dominant limitation.

5 Discussion

The discussion positions the pipeline as a bounded, human-reviewed CLMRS triage layer whose field performance depends heavily on deployment tuning and camera geometry. It also identifies fundamental age-estimation limits, validation gaps, engineering constraints, and governance requirements for protecting children.

Field deployment: 36-fold detection-yield improvement and over-reporting reduction from 9.1 times to as low as 1.8 came from threshold, fusion, and identity-consolidation tuning.The models remained frozen, while zero false merges across 240 proven-different pairs came from a temporal-logic constraint rather than an embedding.
Operational role: A CLMRS triage layer produces confidence-ranked person records with best-shot evidence, age dispersion, and quality flags for staff verification, dismissal, or action.The system is intended to supply continuous, reviewable child-presence candidates rather than autonomous determinations.
Age-estimation limits: Age is unverifiable on 22-to-38-pixel backlit profile faces, so every filed record must remain a lead for human investigation, never evidence of age.The passage attributes this limitation to neighbouring age bands being visually indistinguishable to humans.
Age-estimation limits: 0–6-year-olds are effectively beyond field face-based age estimation, while ages 13–19 show MAE above 5.4 years internally and 7.4 for the reference cascade on FairFace-Africa.Carried infants may be detected geometrically but yield no usable age signal; carry geometry should route directly to human review.
Governance and safeguards: 0.64 precision at 0.83 recall and 0.90 precision at 0.60 recall had identical F1, so the operating point must be chosen explicitly with the child-protection partner.The passage frames false positives as verification labour and false negatives as children left unprotected.
Deployment priorities: Five of six cameras yielded faces for under two per cent of tracked people, making mounting height and tilt the first deployment lever before further model work.A labelled field validation set is identified as the second priority for separating discovery from estimator bias and converting yield into recall and precision.

6 Conclusion

The paper presents an integrated real-time system for child-labour monitoring and evaluates it from proxy footage through an unattended Zimbabwe field pilot. It also reports training, quantisation, and deployment limitations requiring human investigation rather than treating automated outputs as evidence.

Contribution: The system combines real-time individual child detection, child-grade age estimation, identity persistence, and edge deployment in one child-labour monitoring instrument.The paper states that no published system had previously combined these capabilities.
Proxy evaluation: 122 per cent more unique child candidates were surfaced than by the predecessor on 26.8 hours of demanding proxy footage.The pipeline also found 9.3 times more children under ten, with a markedly higher confidence profile.
Field pilot: 36-fold software tuning improved yield during a seventeen-day unattended Zimbabwe pilot covering 38.7 million frames.The pilot produced one logged error and delivered every report.
Field pilot: 1.8–3.9 times the attendance register was the remaining over-reporting after identity consolidation under a simultaneity veto, down from 9.1 times with zero proven-false merges.A separate consolidation variant was also reported, but the supplied passage truncates its result.
Limitations and safeguards: The study documents catastrophic forgetting, validation mismatch, failed FP16 TorchScript and FP4 quantisation toolchains, and misleading sequential camera-setting tests.It also states that adolescent-band accuracy does not support legal working-age decisions and that filed records are leads for human investigation rather than evidence.

A Per-Video End-to-End Results

The section reports per-video end-to-end results for a mixed corpus of child-labour documentaries, feature films, and control recordings, comparing pipeline v5.0 with v6.1 on identical inputs and hardware.

Corpus composition: The corpus combines investigative documentaries on child labour, feature films, and control material.Titles identify the source recordings.
Per-video results: Table 11 reports unique child candidates per video for the end-to-end study.The results correspond to Section 4.5.
Pipeline comparison: Pipeline v5.0 and v6.1 are compared using identical inputs and hardware.The comparison is presented per video.

B Detector Training Configuration

This section documents the complete training configuration for the multi-task detector and notes that augmentation settings were selected through evolutionary hyperparameter search.

B Detector Training Configuration: Table 12 lists the full training configuration for the multi-task detector.The configuration is associated with Section 3.3.
B Detector Training Configuration: The documented detector is multi-task.
B Detector Training Configuration: Augmentation values were obtained by evolutionary hyperparameter search.

C Age-Model Fine-Tuning Phases

The section summarizes successful fine-tuning phases for both cascade members and explains how validation MAE was matched to each phase’s training scope. It also records failed fine-tuning strategies in the main text.

Successful fine-tuning phases: Table 13 summarizes the successful fine-tuning phases for both cascade members.These phases are described as belonging to Section 3.4.3.
Validation protocol: Validation MAE is measured on the validation distribution matching each phase’s training scope.The validation distribution therefore corresponds to the scope used for that phase’s training.
Failed strategies: Failed strategies included high learning rate, pseudo-labelling, self-distillation, mismatched validation, and tenfold oversampling.The main text describes these unsuccessful approaches separately from the successful phases.

D Reproducibility Note

The research prototype used public frameworks, datasets, codebases, and validation splits, while retaining model and training artifacts internally and documenting key fine-tuning modifications.

Reproducibility Note: Public CerberusDet and MiVOLO resources supported detector training, age-model fine-tuning, and benchmark evaluation, with retained model and configuration artifacts summarized in Appendix B.Age fine-tuning incorporated children-only validation filtering and distributed weighted sampling; evaluations used pycocotools with APPA-Real and FairFace validation splits.