부산대학교 도서관

Recently, the field of action segmentation, which involves dividing a time series of behavioral data into fundamental action units, has attracted increasing attention. Although many deep learning models have been proposed, interpreting how these models prioritize features for their decision making remains challenging. In this paper, we propose a novel method that introduces an elastic feature selection technique into existing action segmentation models. By employing the proposed model to assembly processing data, visualizing the significant features that highlight each worker and process becomes feasible. The study shows how the model switches the features in accordance with the prior work process. It also evaluates the validity of the selected features. This analysis is conducted using multi-dimensional behavioral data generated from cameras, acceleration and angular rate sensors, and voice recordings.