Fitting Cornering Speed Models with One-Class Support Vector Machines

This paper investigates the modelling of cornering speed using road curvature as a predictive variable, which is of interest for advanced driver assistance system (ADAS) applications including eco-driving assistance and curve warning. Such models are common in the driver modelling and human factors literature, yet lack reliable parameter estimation methods, requiring an ad-hoc evaluation of the upper envelope of the data followed by linear regression to that envelope. Considering the space of possible combinations of lateral acceleration and cornering speed, we cast the modelling of cornering speed as an `outlier detection' problem which may be solved using one-class Support Vector Machine (SVM) methods from machine learning. For an existing cornering model, we suggest a fitting method using a specific choice of kernel function in a one-class SVM. As the parameters of the cornering speed model may be recovered from the SVM solution, this provides a more robust and reproducible fitting method for this model of cornering speed than the existing envelope-based approaches. In addition, this gives comparable outlier detection performance to generic SVM methods based on Radial Basis Function (RBF) kernels while reducing training times by a factor of 10, indicating potential for use in adaptive eco-driving assistance systems that require retraining either online or between drives.