Part of the research team from the University of Nottingham had the opportunity to present some research to an academic audience at PerCom 2019. This conference brings together academic works on pervasive computing and communications, with applications including wireless networking, mobile and distributed computing, sensor systems, ambient intelligence, and smart devices. This year, the conference took place in Kyoto, Japan.
The research content presented at PerCom covered a wide range of topics, with papers on data collection methods, data processing, security and machine learning applications across all pervasive computing fields. There were also a number of workshops and technical demonstrations at the conference, including PerAwareCity, where we presented our research among academics working on smart cities and transport applications. There were a number of other papers there related to the wider field around Electric Vehicles and V2G, including methods to use location data to detect what kind of transport a person is using, methods to analyse errors in spatial tracking data among others. We presented a piece of work on methods to characterise vehicle usage behaviours based on diverse data sources.
The rest of this blog is a short summary of the work we presented.
Why do we want to know about vehicle usage?
We’ve seen a huge increase in the availability of location data over the last few years: from on-board navigation systems to drivers’ smartphones, vehicles are no exception to this trend. A wide range of services and applications can use this data to provide personalised services, analyse trends over large populations or optimise systems in real time based on a vehicle’s location. This includes route optimization services, vehicle fleet management, car park management, vehicle-to-grid services and other smart city applications.
Many of the above applications need us to be able to understand the patterns of vehicle use. This is particularly important when we might need to predict where our tracked vehicle might be or whether it will be available to be used at some future time. This introduces the need to understand the behavioural patterns of the vehicle users as much as we can from the vehicle data we collect.
How can we collect vehicle data?
While there is a lot of possibility to collect location data from personal smartphones, GPS trackers or inbuilt navigation systems, some issues arise when we want to collect data from a large number of vehicles. It’s often not feasible to get data in exactly the same way from everyone – some on-board vehicle systems might let us export data, others may not collect location data at all. In wide-scale application, it is valuable to be able to process and compare vehicle tracking data that may arrive in different formats.
We also have the problem of whether we are collecting data from the vehicle itself, or from its users. Smartphone location data is widely available, but doesn’t tell us when the user is actually in their vehicle, unless we have some other context. This can make it difficult to be sure that the vehicle is actually in the place that the user’s data suggests, particularly when a vehicle can be driven by multiple users.
As with any application collecting data from people, vehicle tracking also needs to protect the privacy and data security of its users. It is essential that where we collect sensitive data like location, it is treated properly and never transmitted or stored where we don’t need it.
For learning behavioural patterns, this leads us to look for solutions that will:
- Characterise vehicle usage behaviours
- Deal with data coming in different formats
- Be robust to uncertainties in the data
- Avoid explicitly storing location data if not needed.
One of the ways we can start to examine behavioural patterns is to calculate how likely it is that a vehicle is in use at a given time of day. This can be calculated by looking at past behaviours from location data, but doesn’t require the vehicle’s exact location to be known. If we want to add more detail than just ‘moving/not moving’, location can be stored as a more anonymous category such as ‘Work’ or ‘Home’, rather than storing actual location co-ordinates. In our research, we defined the likelihood of vehicle use over a full 24-hour period as a vehicle’s “Activity Profile”. Below is an example of a 24-hour activity profile for a vehicle used for a work commute, with peaks of activity in the morning and evening.
When calculating the activity profile for a large number of vehicles, we start to see certain patterns emerge. In this piece of research, we concentrated on personally-owned vehicles rather than commercial vehicles. From the driving data we analysed, we found four main archetypes of vehicle use:
- Full-time work commute – regular journeys on weekdays
- Semi-regular use – journeys are less structured, but still fairly predictable
- Low Use – vehicle is less likely to be moving at any time of day
- Shift work commute – similar to full-time, but offset to a different time of day.
Once archetypes have been established, we can start to look for interesting behaviours or trends that belong to each group.
If we encounter a new vehicle user, it can take some time to build up enough data to make a personalised profile of their vehicle use behaviours. With a set of vehicle use archetypes, if the new user can be matched to an existing archetype, this gives a more complete profile to start working from.
There are a number of ways we can match a new user to an archetype: this is a fairly common classification problem. Our research tested several machine learning classification methods, including Dynamic Time Warping, Naïve Bayes, K-Nearest Neighbors and Random Forest based classifiers. We also tested what combination of data measures were the most effective for these classifiers. It was found that the vehicle activity profile combined with the percentage of time spent at a ‘Home’ location and the vehicle’s mean daily number of journeys was the most effective combination of data types to match users back to an archetype. In testing, this combination allowed for a correct identification of the exact archetype of 82% of users, sorting 98% of users into the right general archetype category.
This research forms a small part of the wider efforts around electric vehicle and V2G application. The full research paper “Vehicle Availability Profiling from Diverse Data Sources” is to be made available in the Percom 2019 Workshops Proceedings, when these are made available.
“Blog by Dr Sophie Naylor, Research Fellow as part of the Knowledge Transfer Partnership between ATKearney and the University of Nottingham”