Potential Functions as Superposition of Scoring Functions (PF)#
Description#
The general concept of the PF metric is to define a potential function for each static or dynamic object considered by the metric [Wolf2018]. This includes potentials for lane markings, the road geometry, other vehicles, or, in more urban areas, pedestrians and bicyclists. Once a potential function for each object in the scene, denoted by \(U_i(A, S)\), is chosen, one can apply e.g. gradient descent for a given scene \(S\) to the combined potential function \(U(A, S) = U_1(A, S) + \dots + U_k(A, S)\), where \(A\) is an actor and \(k\) denotes the number of objects. A simple example of how to evaluate this metric for an actor \(A_1\) and a given scene \(S'\) is by inserting the values into \(U\), i.e.
However, methods involving the mentioned gradient descent to assess the criticality can improve precision and also provide a suggestion for criticality-reducing vehicle movement.
Due to the way this metric is defined, almost all properties depend on the specified potential functions. Furthermore, while ethical questions play a role when defining any safety surrogate, it becomes more evident for potential functions, as an active decision making in the definition of the potentials is required.
Properties#
Run-time capability#
Yes
Target values#
None found, also highly dependent on the used potential functions
Subject type#
Any, but requires a potential function for each considered subject type
Scenario type#
Depends on specified potential functions
Inputs#
Potential function for each static/dynamic object in the scene that is supposed to be considered, other inputs depend entirely on said potential functions
Output scale#
\([-\infty, \infty]\), number (negative values are possible if goal locations are defined), ratio scale
Reliability#
Largely depends on the used potential functions
Validity#
Largely depends on the used potential functions; no empirical analysis identified
Sensitivity#
Largely depends on the used potential functions
Specificity#
Largely depends on the used potential functions
Prediction model#
Time window#
Depends on quality of potential functions and reliability of computation of the solution to the potential equation problem
Time mode#
Branching time