Visualizing Worklog Based on Human Working Activity Recognition Using Unsupervised Activity Pattern Encoding.

Area of Science:

• Human-Computer Interaction
• Occupational Health and Safety
• Data Science

Background:

• Wearable motion sensors are used for recognizing complex activities during work to improve productivity.
• Current methods face challenges with time-consuming tasks like data labeling and feature extraction.
• Faster modeling requires utilizing unlabeled motion datasets and understanding complex activity characteristics.

Purpose of the Study:

• To develop a novel working activity recognition method that reduces manual modeling efforts.
• To enable faster implementation of activity recognition in practical work environments.
• To evaluate the proposed method's performance against existing techniques.

Main Methods:

• Unsupervised encoding of motion patterns into "atomic activities".
• Representing working activities as combinations of these atomic activities.
• Integrating additional information like sensor time for enhanced recognition.

Main Results:

• Achieved recognition performance (70.3% macro F-measure) equivalent to conventional hand-crafted feature extraction.
• Performance is comparable to previous methods that used large labeled datasets.
• Successfully visualized daily work processes with 71.2% accuracy.

Conclusions:

• The proposed method offers a viable alternative to time-consuming manual modeling in activity recognition.
• It demonstrates potential for rapid deployment of working activity recognition systems in real-world settings.
• This approach can aid in evaluating and improving worker productivity through accurate activity monitoring.