DeepMap+: Recognizing High-Level Indoor Semantics Using Virtual Features and Samples Based on a Multi-Length Window...

Area of Science:

• Computer Science
• Artificial Intelligence
• Human-Computer Interaction

Background:

• Current indoor semantic recognition relies on smartphone sensing, limiting the detail for map enhancement.
• High-level indoor semantics, crucial for map enrichment, are challenging to recognize with existing methods.

Purpose of the Study:

• To develop DeepMap+, an automated system for recognizing high-level indoor semantics using wrist-worn sensing and complex human activities.
• To improve the richness of indoor semantic recognition for detailed map enhancement.

Main Methods:

• Implemented DeepMap+, a deep learning (DL) system with a multi-length window framework to enrich data.
• Introduced novel methods for increasing virtual features and virtual samples to detect complex hand gestures.
• Collected wrist-worn sensor data during 23 high-level indoor semantic activities in a supermarket setting.

Main Results:

• DeepMap+ demonstrated effective recognition of high-level indoor semantics, including public facilities and functional zones.
• The proposed methods for virtual feature and sample augmentation significantly improved classification accuracy.
• The system successfully utilized complex human activities captured by wrist-worn sensors.

Conclusions:

• DeepMap+ offers an effective solution for automated high-level indoor semantic recognition using wrist-worn sensing.
• The integration of DL, multi-length windows, and data augmentation techniques enhances pattern discovery for complex gestures.
• This approach provides a foundation for richer indoor semantic understanding and improved map applications.