Activity Recognition by Classification with Time Stabilization for the SHL Recognition Challenge
Michael Sloma, Makan Arastuie, Kevin Shuai Xu · 2018
This paper describes our submission (team name: Ideas Lab @ UT) to the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge. The SHL recognition challenge considers the problem of human activity recognition from inertial sensor data collected at 100 Hz from an Android smartphone. Our data analysis pipeline contains three stages: a pre-processing stage, a classification stage, and a time stabilization stage. Our main finding is that performing classification on "raw" data features (i.e. without feature extraction) over extremely short time windows (e.g. 0.1 seconds of data) and then stabilizing the activity predictions over longer time windows (e.g. 15 seconds) results in much higher accuracy than directly performing classification on the longer windows. Our submitted model uses a random forest classifier and attains a mean F1 score over all activities of 0.973 on a 10% hold-out sample of the training data, suggesting that activity recognition can be performed quite accurately on the SHL dataset.