Clustering driving styles via image processing

Abstract

It has become of key interest in the insurance industry to understand and extract information from telematics car driving data. Telematics car driving data of individual car drivers can be summarised in so-called speed–acceleration heatmaps. The aim of this study is to cluster such speed–acceleration heatmaps to different categories by analysing similarities and differences in these heatmaps. Making use of local smoothness properties, we propose to process these heatmaps as RGB images. Clustering can then be achieved by involving supervised information via a transfer learning approach using the pre-trained AlexNet to extract discriminative features. The K-means algorithm is then applied on these extracted discriminative features for clustering. The experiment results in an improvement of heatmap clustering compared to classical approaches.

A representation-learning approach for insurance pricing with images

“…Meanwhile, Wüthrich (2017) illustrate how driving styles can be summarized in an image through velocity-acceleration heatmaps. These heatmaps were further used for risk classification; see, for instance, Gao and Wüthrich (2019), Zhu and Wüthrich (2021). From a claims management perspective, the authors of Doshi et al (2023) predict the cost of repairing a car after an accident when provided an image of a damaged vehicle.…”

Insurance pricing with hierarchically structured data an illustration with a workers' compensation insurance portfolio

“…To construct these tariff classes, we either use supervised or unsupervised learning techniques or a combination of both. For example, Henckaerts et al (2021) developed a tariff structure using tree-based machine learning methods, Gao and Wüthrich (2018) employed clustering techniques to group policyholders with similar driving behavior and Zhu and Wüthrich (2021) combined image classification with clustering techniques to differentiate between driving styles.…”