Ensemble approach for sensor-based human activity recognition

Abstract

This paper discusses in detail our (Team:AISA) ensemble based approach to detect Human Activity for the Sussex-Huawei Locomotion-Transportation (SHL) recognition challenge. The SHL recognition challenge is an open competition wherein the participants are tasked with recognizing 8 different types of activities based on smartphone data collected from multiple positions - Hand, Hips, Torso, Bag. On the magnitude of sensor data, time and frequency domain features were calculated to achieve position independence. To make the model robust, we trained it with a random shuffle of the training and validation data provided. To find the optimal hyper-parameters, we parallely executed randomized search to choose the best performing model from about 200 models. We set aside 30% of this combined dataset for internal testing and the model predicted human activities with an F1-Score of 86% on this test dataset.