A Metric to Quantify the Hazard Avoidance Capability of Vehicles

Abstract

Safety is an important parameter considered during the design of Advanced Driver Assistance Systems and fully autonomous vehicles.
One of the ways to assess the road vehicle's safety is by estimating the likelihood with which the vehicle can react to prevent the danger.
In the presence of an impending collision (hazard), the trajectory planning module in the autonomous vehicle would generate few escape trajectories to avoid the collision.
The escape trajectory is chosen such that it maximises the safety of the vehicle based on certain criteria.
One of these criteria is the vehicle's avoidance capability throughout the trajectory.

This thesis presents an avoidance metric that is constructed using a computational procedure to quantify the avoidance capability of the vehicle in both pure longitudinal (1-D) and, combination of both lateral and longitudinal (2-D) scenario.
The key idea is (a) Propagate forward in time the current state of the host (and the world) using a vehicle model to estimate the host's reachable set of states.
(b) Carefully select a set of samples from the reachable set and repeat the propagation.
(c) At every step, the trajectories that lead to collisions are eliminated.
The ratio of the size of the region spanned by the remaining trajectories to the size of the region spanned by all the trajectories (including those that lead to collision) then constitute the estimate of the host's avoidance capability.

Through simulations on specific use-cases, for a pure longitudinal motion, on comparison with Brake Threat Number (BTN), it was observed that the metric (from the proposed computational procedure) performs very similar to BTN and also takes a low computational time of 380 [ms] for a time horizon of 2.5 [s].
However, in the presence of dynamic obstacles, major differences in performance (such as discontinuities, step-like variation), were observed between the metric and BTN.
In the case of the combination of both lateral and longitudinal motion, the computation time for the proposed procedure was found to be independent of the number of obstacles.
To reason about the accuracy of the approximation of the reachable set for a double integrator model obtained from the proposed procedure, it was compared with the nodes obtained from Rapidly-exploring Random Trees (RRT), an under-approximation, and found that the nodes lie either very close or well within the boundary of the approximation.
However, the computation time for the proposed procedure took around 10.87 [s] which comes as a major drawback.