Comparing Feature Learning Methods for Human Activity Recognition: Performance study in new user scenario

Designing robust features for human activity recognition (HAR) that performs well across a wide range of users is a hard task. Therefore, more attention is being given to feature learning techniques, to automatically learn features from raw data. In this paper, we present a comparison study among feature learning methods for HAR. Using accelerometer data, we compare four methods for feature learning from raw-sensor data (PCA-based, clustering, matrix factorization, and LSTM networks) to the traditional hand-crafted feature extraction method. We focus on the performance degradation when each model is evaluated using a new user. According to our results, features learned with Principal Component Analysis are the more robust to the new user scenario. Our results evidence the importance of evaluation in unseen user, since the performance difference compared to a random split testing is big.