Effects of Activity Recognition Window Size and Time Stabilization in the SHL Recognition Challenge

Abstract

The Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge considers the problem of human activity recognition from inertial sensor data collected at 100 Hz from an Android smartphone. We propose a data analysis pipeline that contains three stages: a pre-processing stage, a classification stage, and a time stabilization stage. We find that performing classification on “raw” data features (i.e. without feature extraction) over extremely short time windows (e.g. 0.1 s of data) and then stabilizing the activity predictions over longer time windows (e.g. 15 s) results in much higher accuracy than directly performing classification on the longer windows when evaluated on a 10% hold-out sample of the training data. However, this finding does not hold on the competition test data, where we find that accuracy drops with decreasing window size. Our submitted model uses a random forest classifier and attains a mean F1 score over all activities of about 0.97 on the hold-out sample, but only about 0.54 on the competition test data, indicating that our model does not generalize well despite the use of a hold-out sample to prevent test set leakage.