Transport mode detection and classification from smartphone sensor data using Convolutional Neural Networks

Abstract

Transportation is a significant component of human lives and understanding how individuals travel is an essential task in many fields. Understanding the modes of transport individuals use can lead to improvements in urban planning, traffic control, human health, and environmental sciences. The goal of transport mode detection and classification is to use smartphone devices as human behavioural sensors, to detect and classify individuals movement continuously. Smartphone devices are suitable for transport mode detection, as they are proliferated in modern societies and contain sensors that are suitable for transport mode detection. These sensors include GPS, accelerometers, gyroscopes, magnetometers, barometers, or microphones. The research in this thesis will focus on transport mode detection and classification using data from motions sensors; accelerometers, gyroscopes, magnetometers, and barometers as they do not contain the sensitive private data that is collected when using GPS or microphones. Currently, there are two approaches in state of the art in transport mode detection. In the first approach, time and frequency domain features are extracted from the signals of the motion sensors and used as input to decision tree or neural network machine learning models. In the second approach, Convolutional Neural Networks extract features by finding spatial relations in the signal data and using these for classification. This thesis investigates the use of Convolutional Neural Networks, as they have shown to outperform models trained using time and frequency domain features extracted from the data in the state of the art research. This research studies the effect of different model architectures on the accuracy of Convolutional Neural Network models when using multiple different sensors as input, as well as focusing on which combinations of sensors produce optimal results. Furthermore, the focus will be evaluating the models on real-world data in order to evaluate the feasibility of deploying applications utilizing transport mode detection. This research compares an optimized model architecture along with preprocessing techniques to state of the art Convolutional Neural Network architectures on realworld data. The best baseline algorithm achieved an overall F1 score of 0.57, while the final optimized achieved an overall F1 score of 0.72 on the testing dataset. The optimal combination of motion sensors is with the accelerometer, gyroscope, and barometer.