Real-Time Unsupervised Classification in Industrial Time-Motion Studies Using 2D Key Points

EasyChair Preprint 13681

Abstract

Time-motion studies (TMS), essential for analyzing and optimizing work processes in industrial environments, have traditionally relied on manual observation and data collection, incurring labor-intensive efforts and potential biases. The advancement of computer vision and machine learning presents opportunities for automating and refining TMS through real-time 2D key point data analysis. However, existing methods often require labeled data for training, which can be a significant bottleneck. To address this challenge, we propose a novel approach leveraging unsupervised learning to classify human motions without requiring any labeling, thereby streamlining the process significantly. This study utilizes the Mediapipe framework to extract human skeletal key points, which are processed by a sequence-to-sequence (seq2seq) autoencoder model. The Encoder component captures key point sequences while the Decoder reconstructs these sequences from a compressed latent space. Subsequently, unsupervised clustering techniques are applied to group similar activities, enhancing action recognition efficiency without manual labeling. This innovative methodology eliminates the dependency on labeled data and paves the way for more efficient and scalable TMS in industrial settings.

Keyphrases: Pose Classification, Time-Motion Studies, pose estimation, real-time processing, unsupervised learning

@booklet{EasyChair:13681,
  author    = {Wetu Vexo and Chawalit Jeenanunta and Sapa Chanyachatchwan and Apinun Tunpan and Nisit Sirimarnkit},
  title     = {Real-Time Unsupervised Classification in Industrial Time-Motion Studies Using 2D Key Points},
  howpublished = {EasyChair Preprint 13681},
  year      = {EasyChair, 2024}}