Predicting driver responses to two hazardous scenarios at signalized intersections

Author(s): Ziraldo, Caren, Oliver

Slidedeck Presentation:

Slide deck link

Abstract:

Background:

Safely responding to human drivers during hazardous scenarios is a challenge for autonomous vehicles (AVs), especially during the transition from conventional to fully automated driving. In complex driving scenarios, like navigating an intersection, human drivers provide few clues as to their future behaviour. Additionally, AVs rely on relatively simplistic assumptions about human decision-making to predict and respond appropriately to human driving behaviours. To build predictive models, however, vehicle manufactures must first identify which features of human driving behaviour can be used to predict a hazard response.

Aims:

In this analysis, supervised machine learning methods are used to identify the most predictive features in a data set containing human drivers’ responses to two hazardous scenarios at signalized intersections.

Methods:

To form the data set, 125 licensed drivers (MAge: 22.9 ± 9.4 yrs, Female: 79, Male: 45, Other: 1) responded to hazardous intersection scenarios using a fixed-base, full-vehicle driving simulator. In each scenario, the traffic light turned yellow as the participant driver approached intersection. Specifically, the light changed while the participant driver was in the “dilemma zone” and was therefore forced to decide whether to stop or proceed through the intersection. If the participant chose to proceed, a left-turning hazard vehicle would enter their path, requiring an emergency response to avoid a collision. These responses were represented by features including speed, acceleration, lane position, and position relative to the hazard vehicle. These features were then applied to train a predictive model which would predict a driver’s decision at the yellow and response to the hazard vehicle based on their intersection-approach behaviour.

Results:

From an initial data set of 60 features, a recursive feature elimination protocol was used to rank the predictive power of each of the input features. This protocol also selected the ideal number of features to keep in the data set. Following feature selection, eight were selected for the model. A K-nearest neighbours clustering algorithm trained on 70% of the input data was able to correctly predict the driver’s actions at the intersection 84% of the time.

Discussion:

The clustering algorithm was used to group the driver’s decisions into three categories of action: stopping at the stop bar, proceeding through the yellow and taking evasive action to avoid a collision, and proceeding through the yellow and getting into a collision. The resulting model was more successful in predicting the decision to stop or go through the yellow light than it was at predicting the driver’s ability to avoid a collision. Specifically, the f1-score for the collision classifications was only 0.55. Likely, the decision to proceed through the yellow light and the evasive maneuver will need to be treated as different outcomes to improve the classification accuracy of both problems.

Conclusions:

To reduce the probability of a collision while navigating complex intersections, AVs should be trained to predict the behaviour of human drivers. Reducing the number of collisions is particularly important for AV manufacturers as users' expectations of safety are much higher for AVs than they are for conventional vehicles.