A Computational Driver Model to Predict Driver Control at Unsignalised Intersections

Abstract

The number of cyclists fatally struck when crossing a driver's travel path at an unsignalised intersection has been stable in recent years, indicating that more effort should be made to improve safety in this specific conflict scenario. The most recent safety systems help drivers avoid collisions with cyclists, but improving cyclist safety further requires resolving challenges unique to bicycles and cyclists. In this paper we propose a predictive computational model of driver behaviour in the intersection scenario. Although a handful of studies have focused on describing driver behaviour in this scenario, no computational model that can predict driver control can be found in the literature. The proposed model is based on a biofidelic human sensorimotor-control modelling framework. Two visual cues were used: 1) optical longitudinal looming and 2) projected post-encroachment time between the bicycle and the car. The model was optimised using data from a test-track study in which participants were asked to drive through an intersection where a cyclist would cross their travel path. The performances of the model were evaluated by comparing the simulated driver-control process with the observed control behaviour for each trial using a leave-one-out cross-validation process. The results show that the model performed rather well, reproducing braking controls and kinematics that were similar to the observations. The extent to which the model could be used by safety systems' threatassessment algorithms is discussed. Future research to improve the model's performances is suggested.

Variable-drift diffusion models of pedestrian road-crossing decisions

“…As mentioned in the introduction, most evidence accumulation modeling work in the literature focuses on abstract tasks of perceptual discrimination or value-based choice. Our results instead add to a more limited (but growing) body of support for drift diffusion type models in locomotion and general sensorimotor interaction with the world [2,24,31,41,39,48,71,73]. In this type of context, decisions are less purely "perceptual"or "cognitive", and instead arguably more embodied in nature, yet interestingly the same type of decision mechanisms seem to apply.…”

“…We and others have investigated the application of drift diffusion-type models in the road traffic context, with promising results initially for low-level locomotion decisions on applying braking or steering control [39,48,71], more recently also extending to multi-agent interaction situations [2,24,31,41,73].…”

Accumulation of continuously time-varying sensory evidence constrains neural and behavioral responses in human collision threat detection

“…Third, we show that established evidence accumulation models of decision-making can be extended beyond typical laboratory paradigms with static or intermittently changing abstract stimuli, to a task with clear real-world relevance, and continuously time-varying sensory evidence. We and others have reported that evidence accumulation models show promise for modelling decisions in real-world tasks, e.g., when to apply brakes in response to a developing collision threat [52], [53], [61], or on whether and when to cross a road with oncoming traffic [62], [63]. However, in these types of naturalistic tasks it has not been possible to fit full response time probability distributions per participant, a minimum expectation in evidence accumulation modelling of more typical, abstract laboratory tasks.…”

“…However, computational modelling of evidence accumulation decision-making has so far focused on laboratory paradigms using stimuli that (i) have stationary or only intermittently and/or noisily changing saliency over time [46]- [49], and (ii) are abstract in nature, not mapping directly to any real-world task. We and others have begun exploring the applicability of evidence accumulation models in more naturalistic tasks, particularly in vehicle driving [50]- [53], but so far a conclusive test of these models in such contexts has been lacking: Model fitting is rendered more challenging both because real-world sensory evidence tends to be continuously time-varying, which precludes analytical expression of model likelihood functions, and because of difficulties in obtaining sufficient numbers of repeated trials in ecologically relevant experimental conditions.…”

Modeling Drivers’ Strategy When Overtaking Cyclists in the Presence of Oncoming Traffic

“…Future work may investigate the benefit of Bayesian models to provide richer and arguably more valid information about the driver's uncertainty [58], [59], and biologically inspired models [60]- [62]. Once larger data sets of cyclist-overtaking maneuvers become available, more complex machine learning methods based on neural networks or Markov processes may become feasible alternatives to improve predictive accuracy.…”