Locomotion-Transportation Recognition via LSTM and GPS Derived Feature Engineering from Cell Phone Data

Abstract

This paper was put forth to test the notion of detecting forms of locomotion from various radio frequency data for the 2021 SHL recognition challenge. The model for team Seahawks was created to determine one of 8 different modes of locomotion from the testing dataset which contained data from GPS, WiFi, and cell tower signal data. We found the GPS location to be the most important data especially in combination with some feature engineering to produce the direction and speed of the cell phone in question. These new features were added to the dataset which was subsequently fed into an LSTM based model for evaluation and classification as the dataset was in a time-series format. This resulted in an overall accuracy of.89 on the validation set. This points to the likelihood that the detection of mode of transportation from radio frequency signals of a cell phone is a definite possibility and could be achieved via deep learning and other ML methods.