Using real road driving data to calibrate a model of front-end collision risk

Deployment of first generation collision avoidance systems offers a potential source of abundant data that can be analyzed to assess the dependence of front-end collision risk on speed and headway distance. Passive persistent monitoring of moving cars via wireless telemetry could produce large amounts of realistic data to refine collision warning and auto-braking algorithms. Our objective is to estimate an instantaneous risk function that gives the probability of hard braking conditioned on sudden deceleration to avoid collision with the vehicle ahead. We develop an abstract stochastic model in which this risk depends on two predictors, speed and time-tocollision (TTC), having the same functional form inferred from controlled tests but with parameters that may vary between drivers, roads and traffic environments. The abstract model provides a basis for re-estimating the instantaneous risk by two approaches: (1) Fit a proportional hazards model to the times between hard braking events with speed and TTC as covariates. (2) Minimize the dissimilarity of observed and expected distributions of the retrospective risk, which is instantaneous risk sampled only at hard braking times. Both approaches appear to require an iterative procedure for adaptive parameter estimation via global nonlinear function optimization. One such procedure, illustrated for the second approach with a chi-squared dissimilarity measure, typically attains 5% accuracy in parameter estimation given simulated data describing a few thousand hard braking events.