Loading
How the SCM simulates the full cognitive chain of a human driver from gaze to action.
The Stochastic Cognitive Model is based on a cognitive architecture, meaning that cognitive processes of a human driver can be simulated with the SCM. These cognitive processes are gaze behaviour, the processing of the data in the mental model, situation recognition based on the data and the decision to carry out an action based on the recognised situation.
The SCM is built on a cognitive architecture that replicates each step a human driver performs sequentially and stochastically.
Stochastic gaze distribution matrices determine where and for how long the driver looks, driven by top-down and bottom-up attention.
All perceived information is stored in an internal environment model incomplete and inaccurate, just like real human memory.
Features extracted from the mental model are combined stochastically to recognise the current traffic situation.
Based on the recognised situation, the driver chooses a longitudinal and lateral action executed through vehicle dynamics.
As most of all relevant information during driving is perceived visually, the modelling of human visual perception constitutes an essential part of the Stochastic Cognitive Model. The realistic simulation of human gaze behaviour is highly relevant because all subsequent cognitive processes – including situation assessment and reaction to the given situation – are highly dependent on the information that comes from the gaze behaviour. This is why visual attention modelling is key to traffic interaction modelling.
Gaze behaviour in the SCM is modelled through stochastic processes by applying gaze distribution matrices which specify frequency of gaze allocation and gaze duration. Gaze directions are additionally influenced by top-down or bottom-up attention. Top-down is the voluntary allocation of attention according to the driver's intentions and goals, whereas bottom-up attention is the reflexive allocation of the attention to salient stimuli in the environment. Data as input for the models has been recorded in naturalistic driving studies and simulator experiments.
Bottom-up gaze behaviour represents a higher probability of gaze relocation due to optical stimuli in the surroundings. In the simulation, this is triggered when another vehicle overtakes the ego vehicle, for example, or when another vehicle in the area in front activates the indicator. The video shows the relocation of the SCM agent's gaze target to the front left-hand side when being overtaken by another vehicle. Top-down gaze behaviour is conscious gaze relocation due to driver purposes and intentions. In the simulation, this is especially relevant when it comes to lane change manoeuvres. The video shows the relocation of the SCM agent's gaze target to the left before starting a lane change manoeuvre. Based on the gaze behaviour of the SCM, it is even possible to model distraction scenarios.
All the information that is gathered through the gaze behaviour is handled in the agent's mental model, which represents the driver's internal environment model. This internal environment model is essential for all decision-making processes and cognitive evaluations. Due to the gaze behaviour of the agent and the inherent error-proneness of perception, the information stored in the mental model might be incomplete or inaccurate.
Information is generally classified into three types:
The mental model is not only used to store information but also calculates relevant data for decisions on required actions. For instance, if an SCM agent prepares for a merging manoeuvre, drawing on the perceived data of the other vehicles such as velocities and accelerations the mental model evaluates from the position of the surrounding vehicles if there is enough space for the ego vehicle to merge into a specific gap.
See here for more information on how the agent's mental model can be visualised in simulations.
The SCM agent regularly evaluates whether a change needs to be made to the current actions on the basis of the recognised situation. The decision is made separately through longitudinal and lateral guidance. The lateral guidance is defined by selecting a lateral action state and the longitudinal guidance is defined by selecting a longitudinal action state. Both states only describe the respective action pattern. The detailed execution of the actions in the sense of vehicle dynamics control or the secondary driving tasks is performed in a later step.
Lateral action states can be divided into lane keeping and lane change. Lane keeping includes the intent to change lanes and preparation to merge states for both directions. Lane change actions contain lane changes themselves as well as urgent and comfort swerving acts for both directions. Longitudinal action states differ in free driving, following driving and target braking, covering states such as speed adjustment, approaching, following at desired distance and braking to end of lane.
A central mechanism for critical situations is the evaluation whether a collision with another object can be prevented by braking, evading, both actions or if it is unavoidable. The model uses dynamically calculated TTC thresholds derived from basic kinematic calculations and the Tau Theory. Upper and lower thresholds are implemented to account for uncertainty in human decision-making, which can be prone to misjudgements. The uncertainty is considered by divided intensities between the possible actions using linear interpolation between the threshold values.
Swerving constitutes an evading manoeuvre which does not necessarily require a complete lane change. For an emergency swerve, the possibilities to swerve around an obstacle on the left or on the right are evaluated, including checks for lane boundary crossings and potential side collisions with vehicles on neighbouring lanes. Comfort swerving constitutes a similar manoeuvre but planned in advance and entirely inside the agent's own lane, using the lateral comfort acceleration instead of the maximum and a greater lateral safety distance.
Situation recognition is a core concept in cognitive driver modeling, describing how a driver continuously monitors and interprets the surrounding traffic environment to determine whether their current actions need to change. It involves detecting the behaviour of other road users, identifying potential threats and assessing their relevance to one's own vehicle, all filtered through the limits of human attention and perception.
In the SCM model, this process is formalised through predefined situation patterns organised into microscopic clusters (FRONT, LEFT, RIGHT, REAR — tracking individual vehicles via specific Areas of Interest) and a mesoscopic cluster (covering broader traffic states like traffic jams, blocked lanes or approaching motorway exits). Each recognised pattern receives a base intensity — calculated from the probability of occurrence multiplied by the severity of the required acceleration response, rated on a 0–8 risk scale — so that the most safety-critical situations are prioritised.
To reflect the bounded processing capacity of a human driver, only the top-ranked patterns per cluster are retained, and a new situational decision is only made when the reaction time has elapsed and the pattern ranking has meaningfully changed; two fallback states — FREE_DRIVING (no threat) and COLLISION (post-accident freeze) — ensure the system always has a valid situational output.
The following videos illustrate how the agent's mental model is visualised during simulation runs.
SCM Visualization of the Agent's Mental Model I
One of the most exciting features of the Stochastic Cognitive Model is its ability to visualise the agent's mental model. This feature not only captures what the agent is seeing, but also how the agent is interpreting other vehicles, thereby extrapolating vehicles' movements when they are not visually perceived by the agent. In our visualisation, this is represented by transparent bounding boxes of the vehicles that currently exist in the agent's mental model.
SCM Visualization of the Agent's Mental Model II
With the agent's mental model, the SCM is capable of simulating visual attention and inattention, thus covering a wide range of possible driving scenarios. Visual attention modelling is key to realistic simulated driving scenarios, as it provides essential input for sequential cognitive processes.