engleski

New Dataset for Action Segmentation in Manufacturing Assembly Process

Action segmentation is a fundamental task in many real-world applications of computer vision such as video surveillance, healthcare and human-robot interaction. Although researchers emphasize the need to apply computer vision techniques for solving action segmentation tasks in the context of monitoring and tracking people in manufacturing processes, there is a lack of research focused on this problem. One possible reason for the small number of scientific papers aimed at solving the action segmentation problem in manufacturing processes is the lack of adequate datasets that enable the development and testing of new approaches. We present a new video dataset collected from a manufacturing process consisting of nine different work activities, recorded from two different camera viewpoints, with activities performed by four different workers on three different types of products. In addition to the dataset, we propose metrics for performance evaluation and model comparison. We also present baseline results obtained on proposed metrics.

manufacturing process, computer vision, action segmentation, action detection, deep learning

nije evidentirano