The development of a novel method to identify and describe driving events using only MEMS-sensors in an unmounted smartphone.

Date
2017-03
Journal Title
Journal ISSN
Volume Title
Publisher
Stellenbosch : Stellenbosch University
Abstract
ENGLISH ABSTRACT: The field of vehicle telematics has been revolutionised by mobile- and sensor technology. Mobile phones have become pervasive and the increase in processing capacity, connectivity and richness of sensors have lead to dedicated vehicle telematics hardware being replaced by the widely obtainable and affordable smartphone. The main inconveniences of using mobile phones for telematics sensing are batteryhungry systems (such as GPS) and app-related constraints on device position during operation (as required by systems using Microelectromechanical Systems (MEMS) sensors). Since users are in control of smartphones' permissions and application life cycles, these can be revoked and deactivated at will. Thus, to be a viable solution, the inconvenience caused to the user has to be minimised. This thesis analyses and compares the efficacy and challenges of using the Global Positioning System (GPS) vs. Microelectromechanical Systems (MEMS) sensors. The main metric considered, apart from user convenience, is the ability to detect driving events. It is hypothesised that a generalised sensing platform can be developed to identify a complete and representative set of driving events by exclusively using MEMS sensors in an unmounted smartphone. A system is developed for collecting, annotating and visualising driving data. More than 5000 user-identified driving events' data were collected and labelled using a developed, automated, Fuzzy Logic-based annotation algorithm. MEMS sensor errors and errors induced by sampling from a smartphone-based platform were characterised and a thorough spectral analysis identified high frequency information in MEMS sensor data with specific relevance to driving event detection. A unique multirate processing pipeline was developed using the acquired knowledge. By mitigating the identified errors, exploiting high frequency characteristics in the data and implementing a novel reorientation algorithm, the processing pipeline allows data from an unmounted smartphone to be transformed causally and efficiently into data apt for machine learning. Hidden Markov Models and Random Forests are trained, tested and compared using the processed data. The results of the, better performing, Random Forests show a Balanced accuracy of 70.3% for classifying a complete set of 9 driving events at 2Hz and balanced accuracies of 93%, 88% and 78% for most prevalent events, turning, stationarity and coasting respectively. Though the developed processing and classification systems have not been optimised, or implemented in the form of a smartphone application, the road has been paved for doing so. Given an application that can classify all significant types of driving events in a non-invasive and convenient way, information of specific relevance to an event of interest (i.e. yaw rate for a turn) could be parametrised and uploaded to the cloud. This could have a disruptive effect on the fields of vehicle telematics, intelligent transportation systems and participatory data aggregation.
AFRIKAANSE OPSOMMING: Die voertuig telematika vakgebied is revolutionêr verander deur selfoon- en sensor tegnologië. Selfone is vandag 'n alomteenwoordige verskynsel en 'n beduidende toename in die verwerkingskapasiteit, kommunikasiemediums en sensors-opsies het daartoe gelei dat toegewyde voertuig telematika hardeware, al meer vervang word deur slimfone. Merkbare ongeriewe wat die gebruik van slimfone vir voertuig telematika veroorsaak, is: die batterylewe wat verkort word deur hoë drywing stelsels (soos Globale Posisioneringstelsel (GPS)); en beperkings op die toestel se posisie tydens die toep se werking (soos vereis deur stelsels wat van Mikroelektromeganiese stelsel (MEMS) sensore gebruik). Aangesien slimfoongebruikers volledig in beheer van toepassings se lewenssiklusse is, kan die toep gedeaktiveer word na willekeur. Dus, om 'n werkbare oplossing te word, moet die ongerief van slimfoonafhanklike voertuigtelematikatoepgebruik, tot die minimum beperk word. Dit word in hierdie tesis gepostuleer dat 'n sagtewareplatform ontwikkel kan word, wat 'n volledige en verteenwoordigende stel bestuursgebeure kan identifiseer, deur uitsluitlik MEMS-sensore in 'n ongemonteerde slimfoon te gebruik. 'n Stelsel is ontwikkel vir die versameling, verwerking en visualisering van bestuursdata. Data van meer as 5000 - passasier geïdentifiseerde-bestuursgebeure is ingesamel en met behulp van 'n ontwikkelde, Fuzzy Logic-gebaseerde annoteringsalgoritme, van etikette voorsien. Eienskappe van die MEMS-sensor foute wat veroorsaak word deur die slimfoon monsterproses is geïdentifiseer en maniere is gevind om dit teen te werk. 'n Deeglike spektrale-ontleding het hoëfrekwensie-inligting opgespoor wat aangewend kan word om bestuursgebeurtenisse uit te ken. Met behulp van die verworwe kennis, is 'n unieke multitempoverwerkingspyplyn ontwikkel. Deur die geïdentifiseerde foute teen te werk, hoëfrekwensie-eienskappe te benut en die implementering van 'n unieke heroriënteringsalgoritme, kan die multitempoverwerkingspyplyn, data vanaf 'n ongemonteerde slimfoon kousaal en doeltreffend omskep na data gepas vir masjienleer. Hidden Markov Models en Random Forests is opgelei, getoets en vergelyk met behulp van die ingesamelde-, geannoteerde- en verwerkte data. Die resultate van die toppresteerder, Random Forests, toon 'n Balanced Accuracy van 70.3 % vir die klassi- fikasie van 'n volledige stel van 9 bestuursgebeure teen 2Hz. Hoewel die verwerking en klassifiseringstelsels nie geoptimeer is, of in die vorm van 'n slimfoontoep geïmplementeer is nie, is die fondasie wel neergelê om dit in die toekoms aan te pak. 'n Toep wat alle betekenisvolle tipes bestuursgebeure, op 'n gerieflike manier kan identifiseer, kan inligting wat van belang is vir elke bestuursgebeutenis ontgin en oplaai na die Cloud. Dit kan 'n dramatiese verbetering in velde van voertuigtelematika, intelligente vervoerstelsels en vrywillige dataversameling tot gevolg hê.
Description
Thesis (MEng)--Stellenbosch University, 2017.
Keywords
UCTD, Sensor networks, Smartphones, Automotive telematics, Microelectromechanical systems
Citation