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

Mark Nowak

cs.CVAug 14, 2026

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

Authors: Mark Nowak

Organizations: Conflux Laboratory

Abstract

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 [email protected] 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.

Figures & tables

Panel of Figure 1 — Figure 1: Processing pipeline. Detections feed the age cascade; tracking and re-identification merge frame-level evidence into unique persons; fusion produces one reviewable record per person. Field deployments add a daily aggregation and identity-consolidation layer (Section 3.6 ).

Class	Metric	YOLOv8x-pf	YOLO26x-pf	$\Delta$ rel.
all	[email protected]	0.259	0.415	$+60.3\%$
all	[email protected]:0.95	0.131	0.257	$+96.9\%$
all	AR@100	0.339	0.513	$+51.3\%$
person	[email protected]	0.390	0.683	$+75.0\%$
person	[email protected]	0.162	0.460	$+184.6\%$
person	[email protected]:0.95	0.191	0.438	$+128.8\%$

Table 1: Detector comparison on 13,537 validation images (pycocotools COCO evaluation, identical ground truth). AR@100 denotes average recall at 100 detections per image.

	APPA-Real			FairFace-Africa
Model	MAE	CS@5	CS@10	MAE	CS@5	CS@10
Cascade, FP8 TRT (reference)	4.76	62.6%	90.4%	5.26	60.4%	86.1%
Cascade, FP32	4.75	62.7%	90.5%	5.26	60.5%	86.2%
Cascade, fine-tuned general	5.19	57.9%	87.9%	6.38	51.7%	79.4%
MiVOLO v2, single	4.63	64.3%	90.9%	4.90	63.8%	87.7%
InsightFace (buffalo_l)	13.67	22.1%	44.5%	13.51	23.4%	43.6%

Table 2: Age estimation on APPA-Real (n=1,500) and FairFace-Africa (n=6,165). CS@5/CS@10: share of predictions within 5/10 years. Best value per column in bold.

	APPA-Real		FairFace-Africa
Model	child MAE	child CS@5	child MAE	child CS@5
Cascade, FP8 TRT (reference)	4.74	62.8%	4.64	65.5%
Cascade, FP32	4.75	62.8%	4.64	65.6%
Cascade, fine-tuned general	5.10	61.3%	6.80	51.3%
MiVOLO v2, single	4.72	64.4%	4.17	73.0%
InsightFace (buffalo_l)	22.93	3.4%	18.74	5.6%

Table 3: Child subsets (age

\leq 16

) of the same benchmarks: APPA-Real (n=261) and FairFace-Africa (n=2,736).

Age bucket	n	Cascade FP8	MiVOLO v2 single
0–4	199	2.08	2.66
5–12	1,356	2.63	2.94
13–19	1,181	7.37	5.83
20–39	2,378	5.60	5.21
40–59	880	6.08	6.07
60+	171	6.28	6.28

Table 4: Per-bucket MAE of the reference cascade (FP8) and the single MiVOLO v2 on FairFace-Africa. Buckets with few samples on APPA-Real behave analogously.

Age group	n	MAE	CS@5	Routed to child model
0–3 (infants)	3,162	1.925	94.1%	98.5%
4–7	5,504	2.296	94.6%	97.2%
8–12	3,389	3.211	83.9%	92.4%
13–16	5,315	5.430	50.4%	66.2%
17–18	3,501	5.839	49.3%	40.7%
19–30	45,753	3.956	71.9%	7.6%

Table 5: Cascade evaluation on 146,735 validation samples from fourteen datasets (routing threshold 20 years).

Metric	FP32	FP8 TRT
MAE (years)	4.755	4.757
Median AE (years)	3.725	3.719
CS@5	62.9%	62.6%
CS@10	90.5%	90.4%
Latency, mean (ms)	22.1	12.5
Latency, p50 (ms)	19.5	10.8

Table 6: Age cascade, FP32 PyTorch versus FP8 TensorRT on APPA-Real (batch size 1, NVIDIA GB10).

Metric	FP16 engine	FP8 engine	$\Delta$
Unique child candidates	615	602	$-2.1\%$
Persons tracked	40,363	40,047	$-0.8\%$
Total processing time (min)	959.6	810.5	$-15.5\%$
Mean throughput (FPS)	8.9	10.5	$+18.0\%$
Mean real-time ratio	1.78 $\times$	2.11 $\times$	$+18.5\%$
Mean fusion score	0.843	0.832	$-1.3\%$

Table 7: Full-pipeline ablation of detector precision on the 19-video corpus (482,486 frames); all other components fixed.

Metric	v5.0	v6.1
Unique child candidates	285	634 ( $+122\%$ )
Persons tracked	36,773	40,361 ( $+10\%$ )
Child candidates under age 10	23	213 (9.3 $\times$ )
Minimum fused age (years)	6.3	0.8
Mean fusion score	0.774	0.852
High-confidence share ( $\geq 0.8$ )	49%	69%

Table 8: End-to-end comparison on 19 videos (26.8 h, 482,486 frames). The corpus carries no ground-truth annotations; figures describe system yield, not measured recall.

Metric	Value
Ten-minute segments processed	13,721
Frames processed	38,748,542
Person tracks accumulated	506,672
Segment-level child detections	3,620
Daily evidence packages produced	938
Evidence held for review by the accuracy gate	48

Table 9: Field pilot volume over seventeen days (six cameras, one Jetson AGX Thor).

Window	Encounters	Identities filed	Children on register	Over-report
31 July	77	44	18	$\times$ 2.4
1 August	359	164	18	$\times$ 9.1
2 August	296	141	19	$\times$ 7.4
3 August	173	104	18	$\times$ 5.8
4 August	202	64	18	$\times$ 3.6
5 August	225	63	18	$\times$ 3.5

Table 10: Daily unique child identities filed versus the attendance register (fourteen register-covered windows). Identity consolidation (Section 3.6 ) deployed on 4 August.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Video	Duration	v5.0	v6.1	$\Delta$
Slavery: A Global Investigation	1:17:48	6	66	$+60$
The Gods Must Be Crazy	1:48:50	3	48	$+45$
Shady Chocolate	0:44:59	3	47	$+44$
The Dark Truth of Chocolate	0:22:28	6	34	$+28$
FLA child-labour mitigation project	0:26:22	9	31	$+22$
The Chocolate War	0:57:46	6	25	$+19$

Appendix

Table 11: Unique child candidates per video, pipeline v5.0 versus v6.1 (identical inputs and hardware).

Parameter	Value
Hardware	4 $\times$ NVIDIA H200, DDP, SyncBN, AMP
Input size	$640\times 640$
Batch size	192 (48 per GPU)
Epochs	166 of 300 (early stop; best epoch 115; patience 50)
Wall time	71.7 h
Optimiser	SGD, lr 0 0.00309, one-cycle to 0.0956 $\times$ lr 0

Appendix

Table 12: YOLO26x-pf (CerberusDet) training configuration.

Model	Phase	Epochs	LR	Scope	Val. MAE
General	warm-up (frozen backbone)	5	$3\times 10^{-5}$	all ages	5.039 $\to$ 4.785
General	clean fine-tune	40	$2\times 10^{-6}$	8 datasets	$\to$ 3.729
General	diverse fine-tune	30	$1\times 10^{-6}$	13 datasets	$\to$ 3.936
Specialist	warm-up (frozen backbone)	5	$5\times 10^{-5}$	ages 0–18	3.485 $\to$ 3.050
Specialist	children fine-tune	40	$5\times 10^{-6}$	ages 0–18	$\to$ 2.291
Specialist	infant focus	15	$1\times 10^{-6}$	ages 0–12	$\to$ 1.944

Appendix

Table 13: Fine-tuning phases of the general model and the child specialist.

AI and Labor

AI and labor studies how artificial intelligence changes jobs, wages, skills, hiring, and the organization and quality of work.

See all 95 papers

KidsNanny: A Two-Stage Multimodal Content Moderation Pipeline Integrating Visual Classification, Object Detection, OCR, and Contextual Reasoning for Child Safety
Mar 17, 2026Viraj Panchal, Tanmay Talsaniya, Parag Patel +1Multimodal Content Moderation Content Moderation
KidsNanny Research Team, Vartit Technology Inc.
Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models
Jun 2, 2026Glenn Jocher, Jing Qiu, Mengyu Liu +3Open-Vocabulary Object Detection Instance Segmentation
Ultralytics
Synthetic-Child: An AIGC-Based Synthetic Data Pipeline for Privacy-Preserving Child Posture Estimation
Mar 3, 2026Taowen ZengHuman Pose Estimation Synthetic Data Generation
Independent Researcher

Mar 17, 2026J/K move · Enter open · S save

KidsNanny: A Two-Stage Multimodal Content Moderation Pipeline Integrating Visual Classification, Object Detection, OCR, and Contextual Reasoning for Child Safety

Viraj Panchal, Tanmay Talsaniya, Parag Patel, Meet Patel

KidsNanny Research Team, Vartit Technology Inc.

Multimodal Content Moderation Content Moderation

Open paper AskPDF arXiv

We present KidsNanny, a two-stage multimodal content moderation architecture for child safety. Stage 1 combines a vision transformer (ViT) with an object detector for visual screening (11.7 ms); outputs are routed as text not raw pixels to Stage 2, which applies OCR and a text based 7B language model for contextual reasoning (120 ms total pipeline). We evaluate on the UnsafeBench Sexual category (1,054 images) under two regimes: vision-only, isolating Stage 1, and multimodal, evaluating the full Stage 1+2 pipeline. Stage 1 achieves 80.27% accuracy and 85.39% F1 at 11.7 ms; vision-only baselines range from 59.01% to 77.04% accuracy. The full pipeline achieves 81.40% accuracy and 86.16% F1 at 120 ms, compared to ShieldGemma-2 (64.80% accuracy, 1,136 ms) and LlavaGuard (80.36% accuracy, 4,138 ms). To evaluate text-awareness, we filter two subsets: a text+visual subset (257 images) and a text-only subset (44 images where safety depends primarily on embedded text). On text-only images, KidsNanny achieves 100% recall (25/25 positives; small sample) and 75.76% precision; ShieldGemma-2 achieves 84% recall and 60% precision at 1,136 ms. Results suggest that dedicated OCR-based reasoning may offer recall-precision advantages on text-embedded threats at lower latency, though the small text-only subset limits generalizability. By documenting this architecture and evaluation methodology, we aim to contribute to the broader research effort on efficient multimodal content moderation for child safety.

Abstract

AI and Labor

Explore similar work

KidsNanny: A Two-Stage Multimodal Content Moderation Pipeline Integrating Visual Classification, Object Detection, OCR, and Contextual Reasoning for Child Safety

Ultralytics YOLO26: Unified Real-Time End-to-End Vision Models

Synthetic-Child: An AIGC-Based Synthetic Data Pipeline for Privacy-Preserving Child Posture Estimation