Audio-Based Human Activity Recognition
Rethinking how activity data is collected — faster, more scalable, and less tedious.
Traditional human activity recognition (HAR) datasets rely heavily on video recording and manual annotation, which is time-consuming and difficult to scale. Our project explored whether synchronized audio instructions could replace these methods, making data collection faster and more accessible for both researchers and participants.
We designed a pipeline that aligned audio cues with time-series sensor data from 9-axis IMU devices, allowing activities to be labeled in real time without post-processing. I focused on building data preprocessing scripts and structuring the dataset for LSTM-based models, which ultimately achieved higher validation accuracy compared to button-based and standard baseline methods.
Looking back, I think we could’ve explored more robust generalization across different environments and users. Still, this project shifted how I think about data collection — not just optimizing models, but questioning the assumptions behind how data is gathered in the first place.