Applied Machine Learning for Be-In/Be-Out detection for automated public transportation ticketing based on smartphone sensors

Det siste året har vi samlet inn reisedata på iOS og Android på flere steder fra over 100 deltakere. Ved å analysere de forhåndsbehandlede sensordataene var vi i stand til å trene på flere klassifiseringsalgoritmer og oppnådde en nøyaktighet på over 99% på Android-data og 95% på iOS-data. De foreløpige resultatene er veldig lovende, men vi trenger å gjøre ytterligere undersøkelser videre i prosjektet. Vi har blant annet brukt Shapley-verdier for å forstå den lokale funksjonens betydning. Deretter skal vi teste hvordan de ulike modellene presterer på telefonen. I tillegg til dette vil vi utforske overføringslæring for å forbedre nøyaktigheten og modellytelsen ytterligere. Fra et akademisk perspektiv kommer Philippe til å publisere en kort artikkel med tittelen "Usupervised data mining on spatial-temporal passager mobility and survey data during Covid-19" i en workshop på IEEE Big Data BDA. På grunn av personlige årsaker bestemte Philippe seg for å avslutte sin Ph.D og forlate fluxLoop og Høyskolen Kristiania.

The PhD project will attempt to solve the foundation for enabling automatic ticketing (Be-In, Be-Out/ BIBO) in the public transportation industry. This will lead to better user experience as the traveller is automatically debited for the actual trip and enables more advanced and reasonable pricing structures. Also, the obtained data will provide the industry players with valuable insights for further optimalisation. Past research shows that BLE is the most promising technology for automated BIBO detection. However, the smartphone provides more sensor data which could be useful to accurately identify if a person enters or leaves public transport such as data obtained from the smartphone’s accelerometer or magnetometer. Here, the project will make use of machine learning to correctly classify users in-vehicle presence. The candidate will together with the project owner fluxLoop and the public transport provider Ruter, attempt to develop a reliable measure to identify if a person enters or leaves the public transport. The large amounts of available data and the existing infrastructure allow to create a solution for automated ticketing to increase user experience, enable fair pricing models, while at the same time understand travel behaviour to optimize public transport and reduce fraud. The project will address the use of Bluetooth Low Energy (BLE) in combination with multiple sensors and applied machine learning, which will allow for automated in-vehicle detection. This can provide an easy-to-use and environmentally friendly solution for automated public transport ticketing. During the project different supervised machine learning algorithm will be tested to identify the transportation mode, based on smartphone sensors. Hereby, the focus is on adding more data, considering data as a sequence, as well as conducting error analysis to get insights why models fail the classification task. Future work will include even additional sensors for even better accuracy.