Hidden Markov Model-Based Smart Annotation for Benchmark Cyclic Activity Recognition Database Using Wearables.

Area of Science:

• Wearable sensor technology
• Human activity recognition
• Biomechanical analysis

Background:

• Accurate cycle-level analysis from wearables (e.g., step-counting, gait analysis) is hindered by the scarcity of realistic, labeled datasets.
• Manual annotation of such datasets is labor-intensive and time-consuming.

Purpose of the Study:

• To develop and validate a smart annotation pipeline to automate and streamline the creation of labeled datasets for wearable activity monitoring.
• To establish a large, publicly available dataset for benchmarking activity recognition algorithms.

Main Methods:

• A three-pronged smart annotation approach combining edge detection, local cyclicity estimation, and iteratively trained hierarchical hidden Markov models.
• Collection of synchronized data from 5 inertial measurement units (IMUs), pressure insoles, and video from 80 subjects across 12 diverse activities.
• Significant reduction in manual annotation effort, requiring adjustment for only 14% of events (8% for walking-dominant scenarios).

Main Results:

• A novel smart annotation pipeline reducing manual labeling effort to 14% (8% for walking).
• Creation and public release of a large dataset containing over 150,000 labeled cycles from 80 subjects across 12 activities.
• The dataset encompasses various motion types, including steady-state, transitions, and directional changes, using multi-modal sensor data.

Conclusions:

• The proposed smart annotation pipeline effectively reduces the burden of dataset creation for wearable activity monitoring.
• The publicly available dataset and annotation pipeline provide a valuable resource for developing and validating semi- and unsupervised activity recognition algorithms.
• This work establishes a benchmark for future research in realistic human activity analysis using wearable sensors.