Estimating road friction from kinematic summaries at curved sections

Abstract

We present a system for estimating the friction of the pavement surface at any curved road section, by arriving at a consensus estimate, based on data from vehicles that have recently passed through that section. This estimate can help following vehicles. To keep costs down, we depend only on standard automotive sensors, such as the IMU, and sensors for the steering angle and wheel speeds. Our system’s workflow consists of: (i) processing measurements from vehicular sensors, (ii) transmitting short kinematic summaries from vehicles to a road side unit (RSU), using V2X communication, and (iii) estimating the friction coefficients, by running a machine learning regressor at the RSU, on summaries from individual vehicles, and then combining several such estimates.We study two key questions: (i) should each individual road section have a local friction coefficient regressor, or can we use a global regressor that covers all the possible road sections? and (ii) how accurate are the resulting regressor estimates? We test the performance of design variations of our system, using simulations of normal driving scenarios at curved road sections, using the commercial package Dyna4. We consider a single vehicle type with varying levels of tyre wear, and a range of road friction coefficients. We find that: (a) only a marginal loss of accuracy is incurred in using a global regressor as compared to local regressors, (b) the consensus estimate at the RSU has a worst case error of about ten percent, if the combination is based on at least fifty recently passed vehicles, and (c) our regressors have root mean square (RMS) errors that are less than five percent. The RMS error rate of our system is half as that of a commercial friction estimation service [7].But when tested with data from extreme driving manoeuvres that were unseen in the training data, our regressor performs an order of magnitude worse than on data from normal driving runs on curved road sections. Still our regressor’s RMS errors on such test data are no worse than the state of the art Artificial Neural Network regressors [18], [30].