Unsupervised driver performance scoring using geographical position and sensor data

Rating driving performance is a challenging topic. It attracts professionals from a variety of domains such as automotive industry and insurance companies. A great majority of the previous studies combines multiple measurement modalities such as Controller Are Network (CAN Bus) data, physiological measurements, camera reconrdings and localization estimates from Global Positioning System (GPS). One school of thought attempted to discriminate agressive/non-agressive, attentive/inattentive or drowsy/wakeful drivers through a statistical learning. Other researchers applied a rule-based approach. However, this approaches are inapplicable since labelled data for supervised learning schemes is scarce and rules that are representative for all road conditions are not feasible. Moreover, the abundance of sensor modalities in a personal vehicle is rather costly. In order to handle these problems, in this work, we propose a fully unsupervised driving style scoring mechanism operating on a minimalistic dataset. The proposed model operates similar to conventional anomaly detecton schemes. In this setting, a driving experience is scored in proportion to its congruency to the driving norms which are obtained as the most common driving patters in the training data. As a novelity of our work, these norms are defined considering road type and traffic flow patterns. This is applied via a probabilistic approach where joint probability densities of the variables controlling road type, traffic flow type and driving style are required. Since estimating this probability is mathematically intractable, we follow an alternative approach relaxing the probability estimation through discretization. In this context, each of these variables are clustered by unsupervised learning techniques and the joint probabilities are approximated by the number elements shared between inter-variable clusters. This probability information is stored in a special architecture which we call Co-Clustering Matrix. (CCM). Once this matrix is learnt, scoring of a new driving experience is degraded into finding its position inside the matrix. Clustering of these variables is the central point of our work. This part includes clustering of road types through GPS recordings and traffic flow type and driving style clustering by CAN Bus data as well as the identification of the most efficient clustering methods and distance metrics. All evaluations are supported by mathematical evidences and the factors behind successful methods are discuessed. In this way, we attempt to present a framework for the prospective studies. Furthermore, we discover the efficiency of the clustering of lower dimensional representations rather than the original feature sets. Upon obtaining successful clustering of the data from multiple views, we validate our scoring mechanism utilizing a small labelled dataset. Here, the aggressive drivers are expected to obtain significantly lower scores than their nonaggressive counterparts. This is achieved and statistically validated. Following that, we follow the same procedure for another scoring methodology and in contrast to our approach, no change is observed between scoring patterns of aggressive and nonaggressive drivers.