What Actually Works for Activity Recognition in Scenarios with Significant Domain Shift: Lessons Learned from the...

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• The Sussex-Huawei Locomotion-Transportation Recognition Challenge (2018-2021) focused on recognizing human activities using smartphone sensor data.
• Previous challenges explored cross-location and cross-person activity recognition.

Purpose of the Study:

• To analyze the effectiveness of machine learning pipeline components for activity recognition.
• To identify key factors influencing the performance of locomotion and transportation recognition models.

Main Methods:

• Utilized challenge scenarios focusing on cross-location and cross-person recognition.
• Evaluated data selection, model specialization (locomotion vs. transportation), semi-supervised learning, and temporal smoothing (Hidden Markov Models).

Main Results:

• Location-specific data selection improved F1 scores by up to 10 percentage points.
• Temporal smoothing with Hidden Markov Models improved performance by nearly 10 percentage points.
• Separate models for locomotion and vehicle transportation, and semi-supervised learning, offered smaller gains (approx. 1 percentage point each).

Conclusions:

• Appropriate data selection and temporal smoothing are crucial for enhancing activity recognition accuracy.
• The utility of advanced feature selection and clustering for person-specific models requires further case-by-case investigation.