[1Yin-A] Poster 1

[1Yin-A-22] A Study on Zero-Shot Prediction of Repetitive Task Duration in Manufacturing Using an Action Counting Model

〇Kyoka Yoshioka¹, Ryo Morita¹, Masayuki Satou¹, Akihiko Imajo¹, Shunya Oishi¹ (1. KONICA MINOLTA, INC.)

Keywords:

Action Counting,Zero-shot prediction,Periodic Action Recognition,Work Time Prediction,Action Recognition

In the manufacturing industry, accurately measuring each worker’s task duration is essential for improving productivity. Temporal Action Localization (TAL) models, which predict the start and end times of actions within videos, are potential tools for this purpose. However, most TAL models face the limitation of being unable to perform zero-shot inference. In contrast, Action Counting models can estimate the number of repetitive actions in a video in a zero-shot manner by leveraging frame-to-frame similarity using pre-trained models. Although Action Counting models are originally designed for counting repetitive actions, this study focuses on the fact that many manufacturing tasks—such as assembly and polishing—are inherently repetitive. We investigate whether RepNet, a representative Action Counting model, can be used to predict the start and end times of repetitive actions in manufacturing processes. Furthermore, we extend the model to handle non-visual inputs such as sensor data to enhance prediction accuracy. Experimental results demonstrate that our approach achieves a satisfactory level of accuracy, suggesting that Action Counting models have the potential to enable zero-shot measurement of task durations in manufacturing environments.