Pennino, Federico (Alma Mater Studiorum - Università di Bologna), Sette, Davide (Ducati Motor Holding s.p.a.), Attisano, David (Ducati Motor Holding s.p.a.), Gabbrielli, Maurizio (University of Bologna)

Driving Style Representation Via Convolutional Neural Networks: A Contrastive Learning Approach

Scheduled for presentation during the Poster Session "Data Mining and Data Analysis" (FrAT14), Friday, September 27, 2024, 10:30−12:30, Foyer

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 3, 2024

Abstract

Being able to profile the way a user is driving has always been an important task in automotive context. This kind of information can help companies to address their decision on product development and to improve users’ performance. Furthermore, sometimes it is not possible to know a priori who is riding, and it may be necessary to associate particular features with the style itself implemented by the rider. However, even though companies accumulate a big quantity of data, an accurate process of data labelling — a crucial component in classical machine learning — is not always possible. In fact, such a data labelling process usually requires an important effort in terms of time and domain knowledge. This is the reason why being able to perform driving style recognition without the need of accurate label data collection could help companies. We present here a machine learning framework able to recognize the driver style using time series coming from sensors mounted on a motorcycle. We used Contrastive Learning to address this challenge by learning a feature space in which time series coming from the same driver are brought closer together while pushing dissimilar apart. We used then clustering to extract similar patterns – that define our driving style – over the representation space. This solution has been applied on data collected on the road and at a racetrack, showing to be robust at different levels of analysis.