Location-based Human Activity Recognition Using Long-Term Deep Learning Invariant Mapping

Abstract

In the area of human activity recognition based on mobile phone data, the Sussex-Huawei locomotion dataset[6][19] has been gaining popularity over recent years. This dataset contains sensors' readings from several mobile phones that have been collected while their users participated in eight types of their everyday offline activities: staying still (while sitting or standing), walking, running, riding a bike, staying in a car, on a bus, the train or taking an underground. The main goal of this project was to recognize those activities based on human location and wireless signal receivers data (such as Wi-Fi, GPS, LTE, CDMA receivers). This paper proposes using an automatic feature extraction deep learning algorithm to extract long-Term patterns and features that maximize differentiation between activities. We first determined essential data characteristics (mean, standard deviation, min and max values, difference, distance, velocity, etc.) and trained an ensemble of Siamese LSTM feature extractors on training data during experiments. Then we selected the best features using validation data and trained the classifier on training and validation datasets. According to the model evaluation results, as the LPEM team, we achieved 79.89% of the F1 score using 4-fold shuffling.