In the last years, we assisted to a great hype around autonomous vehicles (AVs). However, if we want to see streets where AVs and human drivers live side by side, AVs must be able to dive into our transportation system. For this reason, human-like driving and decision-making processes represent a hot topic in the AVs research field.
Misunderstandings between AVs and human drivers are the cause of accidents that have been taken into account by the Nanyang Technological University of Singapore, which designed a novel human-like driving and decision-making framework for AVs. Since lane change is one of the most frequent types of car accidents, this research paper mainly studied human-like lane-change decision-making for AVs. The architecture of the proposed framework is depicted in the following figure.
First, based on the driving behavior analysis of human drivers, the fuzzy inference approach is used to build an Aggressiveness Estimation model of surrounding vehicles. Then, the Game-Theoretic Decision-Making module makes a lane-change decision taking into account the trade-off between driving safety and travel efficiency, as well as considering the interaction between the AV and the closest vehicle driven by a human as a dynamic game. The lane-change decision is then provided to the human-like BELCM-based Driving Model, which, as we will see, exploits the two-point preview model and the brain emotional learning circuit model (BELCM). Finally, the outputs of the last two models are provided as input to the Vehicle Control module to control the motions of the AV. In the next paragraphs, we are going to further analyze the first three modules of the proposed framework.
Aggressiveness Estimation Model
First of all, the researchers analyzed through a proper dataset the driving behaviors of humans that reflect their aggressiveness. Generally, human drivers’ aggressiveness is caused by the following driving behaviors: accelerating/decelerating and steering. These behaviors modify the velocity and the yaw rate of the vehicle. Based on the distributions of velocity and yaw rate in the considered dataset, the fuzzy inference approach is able to estimate the aggressiveness level of a driver, given the velocity and the yaw rate of her vehicle. In general, higher values of velocity and yaw rate correspond to higher aggressiveness levels.
BELCM-based Human-Like Driving Model
The Brain Emotional Learning Circuit Model (BELCM) is a brain-like computing model that simulates the operating mechanism of the human brain. In particular, here, this model is used to simulate the driving behaviors of human drivers. In general, the BELCM can realize a tracking control system by combining Stimulus Inputs (SIs) and an Emotional Signal (ES). SI and ES are provided to the BELCM-based model thanks to the Two-point Preview Model, illustrated in the following figure:
In the scientific literature, it has been studied human drivers tend to make their decisions based on two points called the near point and the far point, which are shown in the figure through the N and the F, respectively. The idea is to exploit the Two-point Preview Model to generate the SIs and the ES, that will be provided as input to the BELCM-based Human-Like Driving Model. This model then computes the control output to the vehicle, also considering the human’s physical delay between the mental signal processing and the muscular activation.
Game-Theoretic Decision-Making ModuleIn this module, at first, the researchers proposed a collision risk assessment algorithm in order to guarantee lane-change safety. This algorithm takes into account different parameters like driving aggressiveness and some properties (width and length) of the autonomous vehicle and the closest vehicle driven by a human that is going to be affected by the lane-change decision. Then a decision-making cost function is designed to balance the trade-off between driving safety and travel efficiency. Finally, the dynamic game theory is applied to transform the lane-change decision-making process into a dynamic game, where both the autonomous vehicle and the human driver want to minimize their decision-making cost function.