All Issue

2026 Vol.5, Issue 1 Preview Page
Comparative Study on Anomaly Detection Models to Infer Workers’ Walking Hazards using Video Keypoint Detection

CCTV 영상 키포인트 감지 기반의 건설 현장 작업자 보행 위험 추론을 위한 이상치 탐지 모델 비교 연구

Taeyoon Kim1
,
Sungjoo Hwang2*
,
Siyeon Kim3
김 태윤1
,
황 성주2*
,
김 시연3
1Ph.D. Student, Department of Architectural & Urban Systems Engineering, Ewha Womans University
2Corresponding Author, Member, Associate Professor, Department of Architectural & Urban Systems Engineering, Ewha Womans University
3Postdoctoral Research Fellow, School of Construction Management Technology, Purdue University
1이화여자대학교 건축도시시스템공학과 박사과정
2교신저자 ‧ 종신회원 ‧ 이화여자대학교 건축도시시스템공학과 부교수
3퍼듀대학교 건설관리기술학부 박사후 연구원
*Corresponding Author
31 March 2026. pp. 8-16
PDF XML Citation
Abstract

Construction sites are high-risk environments where changing obstacles pose dangers of falls and collisions for workers. Traditional safety management, relying on visual inspections or wearable sensors, faces practical limitations such as equipment costs and worker resistance. To address these issues, this study proposes using existing CCTV infrastructure and video-based keypoint detection technology to detect walking abnormalities. We define the spatiotemporal correlation between the left and right ankle coordinates as a feature and compare the performance of four anomaly detection methods (Z-score, IQR, Mahalanobis Distance, LSTM Autoencoder). The results show that statistical methods often produce false positives due to a lack of temporal context, while the LSTM Autoencoder model performed best with an F1-score of 83.1%. This study demonstrates the feasibility of building an intelligent safety management system that identifies potential walking hazards through video analysis alone, without additional equipment.

Keywords
Construction Safety
Keypoint Detection
Anomaly Detection
Gait Analysis
LSTM Autoencoder

건설 현장은 시시각각 변화하는 비정형적 장애물로 인해 작업자에게 전도, 충돌, 낙상 위험을 초래하는 고위험 환경이다. 기존 육안 점검이나 웨어러블 센서 기반의 안전 관리는 추가 장비 비용과 착용 거부감 등 실무적 한계에 직면해 있다. 이를 극복하기 위해 본 연구는 기존 CCTV 인프라와 영상 기반 키포인트 감지 기술을 활용하여 작업자 안전을 위한 보행 이상을 탐지하는 방법론을 제안한다. 건설현장 대체 환경에서 작업자의 좌우 발목 좌표 간 시공간적 상관관계를 특징으로 정의하고, 통계 및 기계학습을 포함한 4가지 이상치 탐지 기법(Z-score, IQR, Mahalanobis Distance, LSTM Autoencoder)의 성능을 비교 분석하였다. 실험 결과, 통계 기반 기법들은 시계열적 맥락을 반영하지 못해 오검출이 잦았던 반면, LSTM Autoencoder 모델은 F1-score 83.1%로 가장 우수한 성능을 입증하였다. 또한, 본 연구는 별도의 장비 도입 없이 영상 분석만으로 작업자의 신체 반응을 파악하여, 잠재적 보행 위험 구역을 식별하는 지능형 안전 관리 시스템의 구축 가능성을 확인하였다.

키워드
건설 안전
키포인트 감지
이상치 탐지
보행 분석
LSTM 오토인코더
References
1

Cao, Z., Simon, T., Wei, S. E., and Sheikh, Y. (2017). Realtime multi-person 2D pose estimation using part affinity fields. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), IEEE, Honolulu, HI, USA, pp. 7291–7299. https://doi.org/10.1109/CVPR.2017.143

10.1109/CVPR.2017.143
2

Darban, Z. Z., Webb, G. I., Pan, S., Aggarwal, C. C., and Salehi, M. (2022). Deep learning for time series anomaly detection: A survey. arXiv Preprint arXiv:2211.05244. https://doi.org/10.1145/3691338

10.1145/3691338
3

Githinji, S., and Maina, C. (2023). Anomaly detection on time series sensor data using deep LSTM-autoencoder. 2023 IEEE AFRICON, 1–6. https://doi.org/10.1109/AFRICON55910.2023.10293676

10.1109/AFRICON55910.2023.10293676
4

Ibrahim, K., Simpeh, F., and Adebowale, O. J. (2025). Benefits and challenges of wearable safety devices in the construction sector. Smart and Sustainable Built Environment, 14(1), pp. 50–71. https://doi.org/10.1108/SASBE-12-2022-0266

10.1108/SASBE-12-2022-0266
5

Kanu-Asiegbu, A. M., Vasudevan, R., and Du, X. (2021). Leveraging trajectory prediction for pedestrian video anomaly detection. In 2021 IEEE Symposium Series on Computational Intelligence (SSCI) (pp. 01–08). IEEE. https://doi.org/10.1109/SSCI50451.2021.9660004

10.1109/SSCI50451.2021.9660004
6

Kim, S., Kim, T., Ham, Y., and Hwang, S. (2024). Feasibility of anomaly detection for inferring pedestrian discomfort in response to environmental factors using pedestrian footstep data. Available at SSRN 4942717. https://dx.doi.org/10.2139/ssrn.4942717

10.2139/ssrn.4942717
7

Lee, J., Kim, T., Beak, S., Moon, Y., and Jeong, J. (2023). Real-Time Pose Estimation Based on ResNet-50 for Rapid Safety Prevention and Accident Detection for Field Workers. Electronics, 12(16), 3513. https://doi.org/10.3390/electronics12163513

10.3390/electronics12163513
8

Li, X., Du, H., and Wu, X. (2023). Algorithm of Pedestrian Pose Recognition Based on Keypoint Detection. In 2023 8th IEEE International Conference on Network Intelligence and Digital Content (IC-NIDC) (pp. 122–126). IEEE. https://doi.org/10.1109/IC-NIDC59918.2023.10390775

10.1109/IC-NIDC59918.2023.10390775
9

Wang, C., Li, Y., Xiong, Z., Luo, Y., and Cao, Y. (2021). Lower body rehabilitation dataset and model optimization. In 2021 IEEE International Conference on Multimedia and Expo (ICME) (pp. 1–6). IEEE. https://doi.org/10.1109/ICME51207.2021.9428432

10.1109/ICME51207.2021.9428432
Information
  • Publisher :Korean Society of Automation and Robotics in Construction
  • Publisher(Ko) :(사)한국건설자동화·로보틱스학회
  • Journal Title :Journal of Construction Automation and Robotics
  • Journal Title(Ko) :건설자동화·로보틱스 논문집
  • Volume : 5
  • No :1
  • Pages :8-16
  • Received Date : 2026-03-11
  • Revised Date : 2026-03-23
  • Accepted Date : 2026-03-23
  • DOI :https://doi.org/10.55785/JCAR.5.1.8
Journal Informaiton Journal of Construction Automation and Robotics Journal of Construction Automation and Robotics
Online Submission
  • crossref crossmark
  • crossref cited-by
  • orcid
  • open access
  • ccl
Journal Informaiton Journal Informaiton - close