Leveraging Machine Learning and Wearable Cameras to Analyze Children's Social Interactions

Aug 3, 2026

Direct insights into children’s daily experiences are limited despite their importance for development. Traditional lab-based play sessions do not capture naturalistic social behavior, and wearable video creates large volumes that strain manual coding. A machine-learning pipeline is introduced to analyze everyday interactions from 224 hours of camera-wear data collected from 64 children aged 3–5 in Leipzig since 03/2020, focusing on presence, face, gaze, and voice cues. Results use YOLO to reach 80% person detection accuracy and 0.9 F1 for face detection.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages2

Reading Time5 min

AccessFree to download

Motivation and challenge

Limitations of laboratory methods

Data volume from wearable recording

Proposed machine-learning approach

Social interaction cues targeted

Dataset and collection protocol

Model performance and preliminary findings

Person and face detection results

Observed interaction patterns

Next steps

Gaze and voice integration

UC Merced Proceedings of the Annual Meeting of the Cognitive Science Society Title Leveraging Machine Learning and Wearable Cameras to Analyze Children's Social Interactions Permalink [https://escholarship.org/uc/item/4r5284w3](https://escholarship.org/uc/item/4r5284w3) Journal Proceedings of the Annual Meeting of the Cognitive Science Society, 47(0) Authors Suffo, Nele-Pauline Martin, Pierre-Etienne Zahra, Anam et al. Publication Date 2025 Copyright Information This work is made available under the terms of a Creative Commons Attribution License, available at [https://creativecommons.org/licenses/by/4.0/](https://creativecommons.org/licenses/by/4.0/) Peer reviewed [eScholarship.org](eScholarship.org) Powered by the California Digital Library University of California Leveraging Machine Learning and Wearable Cameras to Analyze Children’s Social Interactions Nele-Pauline Suffo Leuphana University, Lüneburg, Germany Pierre-Etienne Martin Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany Anam Zahra Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany Daniel Haun Max Planck Institute for Evolutionary Anthropology, Leipzig, Saxony, Germany Manuel Bohn Max Planck Institute for Evolutionary Anthropology, Leipzig, Germany Abstract Direct insights into children’s daily experiences are limited despite their importance for development (Rogoffet al., 2018) . Traditional methods, like laboratory play sessions, fail to capture naturalistic interactions. Wearable recording devices provide richer data, but their sheer volume challenges traditional coding. We introduce a machine-learning approach to analyze children’s everyday interactions. Sixty-four children (ages 3–5) in Leipzig, Germany, wore vests with small cameras, recording 224 hours of video since 03/2020 . Our analysis focuses on social interaction cues: person presence, face, gaze, and voice. Using YOLO11, we achieved 80% accuracy in person detection and a 0.9 F1 score for face detection. Preliminary results indicate children often spend time alone or with one person, with face presence in only 17.75% of frames. We will next integrate gaze and voice detection to assess child-directed speech. Our machine-learning approach provides novel insights into children’s natural social environments, advancing research on early development. 6423 In D. Barner, N.R. Bramley, A. Ruggeri and C.M. Walker (Eds.), Proceedings of the 47th Annual Conference of the Cognitive Science Society ©2025 the author(s) . This work is licensed under a Creative Commons Attribution 4.0 International License (CC BY) .

FAQs

Why are traditional lab play sessions insufficient for studying children’s social interactions?

Laboratory play sessions do not reflect naturalistic, everyday interactions. This limits insight into real-world social behavior during development.

What data was collected and how was it recorded?

The study involved 64 children aged 3–5 in Leipzig, Germany, wearing vests with small cameras. The dataset covers 224 hours of video collected since 03/2020.

Which social cues does the analysis target, and what were the preliminary detection results?

The approach focuses on presence, face, gaze, and voice cues. Using YOLO, it reports 80% accuracy for person detection and a 0.9 F1 score for face detection.

Aug 3, 2026

Upload by Seraphina