Digital twins predict future AV crash scenarios - D2.4 Summary
SAFE-UP Deliverable 2.4 focuses on the necessary steps in developing a simulation environment that can model future safety-critical scenarios involving fully autonomous vehicles (AVs). A traffic simulation environment attempts to represent what could happen in a scenario that cannot be tested in real-life conditions – either because it is not yet feasible or too costly. In general, the models included in simulation environments are collision-free by design, which is quite the obstacle when studying potential (new) crash scenarios.
That’s why it is necessary to create the conditions necessary in the simulation environment to represent a crash or near-crash situation. The environment will not show actual collisions but will contain information about unsafe user behaviour and identify near-collision scenarios. For SAFE-UP, two types of environments are analysed: urban and non-urban, based on real and synthetic networks. A real network represents the roads from an existing city while a synthetic one is fabricated. The importance here is that the road network used contains all possible scenarios that can be encountered (i.e., one type of each possible intersection, straight and bendy roads, etc.), and all potential users.
The first step in creating the simulation environment is collecting the data. By using data on road geometry, traffic control, transit, and network performance, the digital twin traffic network can be created and used as the foundation for the vehicles to interact on. Information about user behaviour and demand is also added to the model at that stage, before it is calibrated, both for traffic efficiency and safety.
The calibration for traffic efficiency checks that the number of cars in the network matches real scenarios, while the calibration for traffic safety ensures that the number of collisions or near collisions corresponds to reality. The key part of the simulation for the project is to model unsafe driver behaviour.
To do so, new behavioural models were created to bridge the gap between the current modelling practices and actual decision-making processes, as well as a realistic model for AV driving conditions:
The AV model was developed using conditions from onboard AV software.
The human driver model is set to include errors made by the driver(s) by creating a two-layer model. One layer contains the collision-free models and the other incorporates driver perception error.
The pedestrian and cyclist models were created using virtual reality simulations while recording the movement and decisions of test persons.
The power two-wheeler model uses artificial intelligence to model each vehicle as realistically as possible, replicating their decision making, since riders do not always follow traffic lanes.
The calibration and validation steps are vital in ensuring the model corresponds to reality. The calibration compares the results of the models to real-life scenarios using statistical tools, while the validation ensures that the models will correctly adapt to any changes.
The new models previously mentioned will be included in a co-simulation environment to merge all the users into one model. All the models are run and managed by different computers but are transposed in Aimsun Next, the simulation environment. Once the other vehicles are positioned in the network, the vehicles generated by Aimsun Next will react to the other users according to the transport modelling principles (i.e., car following, lane changing and collaboration principles). The ultimate goal is to record unsafe behaviour and near-crash scenarios to predict future safety-critical scenarios that may occur with the roll-out of AVs. To classify the safety of these vehicle interactions, metrics have been developed and reported in D2.5.
Check out our full list of public deliverables on our Resources page.
Want to get in touch with the SAFE-UP team? Email us at: contact@safe-up.eu