Safety

Collisions#

sumo tracks gaps between vehicles that are on the same edge either fully or partially. By default, whenever these gaps are reduced to below a vehicles minGap a collision is registered (default 2.5m). This (potentially) surprising behavior is used to detect issues with the default car-following model that is supposed to never reduce the gap below the minGap. By setting the option --collision.mingap-factor this threshold can be reduced. When setting --collision.mingap-factor 0, only physical collisions (i.e. front and back bumper meet or overlap) are registered.

When setting the option --collision.check-junctions, collisions between vehicles on the same intersection are also checked by detecting overlap of the vehicles shapes.

Note

Junction collisions are never registered on junctions of type unregulated.

The consequence of a collision is configured using the option --collision.action using one of the following keywords:

  • teleport: (the default): the follower vehicle is moved (teleported) to the next edge on its route
  • warn: a warning is issued
  • none: no action is taken
  • remove: both vehicles are removed from the simulation

Additionally there is the possibility of stopping vehicles for a fixed time before the collision action takes place. This allows for pile-ups and traffic disturbance. To enable stopping, the option --collision.stoptime <TIME> must be set with the stopping time in seconds.

Note

The pedestrian model striping detects collisions between pedestrians. This only serves to detect issues with the model.

Deliberately causing collisions#

A simple way to create collisions is by using TraCI to override save speeds (with traci.vehicle.setSpeed) or forcing unsafe lane changes. To create collisions in this way it is also necessary to disable safety checks using the commands speedMode and laneChangeMode.

Alternatively, various models within SUMO may be configured to making driving less safe. This can be used to create collisions with some probability (or with certainty if vehicle stops are used to force unexpected braking in a critical situation).

Collisions during car-following#

Rear-end collisions during normal driving may be caused by any of the following:

  • vehicles with a value of tau that is lower than the simulation step size (default 1s) when using the default Krauss model.
  • vehicles with a value of tau that is lower than their actionStepLength
  • vehicles with an apparentDecel parameter that is lower than their decel parameter (causing other drivers to misjudge their deceleration capabilities)
  • driver imperfection modelled with the driverstate-device

Collisions from lane-changing can be caused by unsafe lateral movements (side collisions) and by changing lanes in a way that creates unsafe following situations (rear-end collisions).

Side collisions can be caused by - configuring lateral imperfection with vType parameter lcSigma - allowing lateral encroachment with vType parameter lcPushy (but this parameter itself will not cause collisions, only reduce lateral gaps in some situations, requires the sublane model) - lcImpatience (growing impatience permits lower lateral gaps when using the sublane model)

Unsafe changing can be caused by configuringlower gap acceptance with parameter - lcAssertive (the acceptable gap is computed by dividing the safe gap by lcAssertive)

Collisions at Intersections#

Collisions at intersections may be caused by any of the following:

  • Unsafe junction model parameters
    • jmDriveAfterRedTime > 0 (ignoring a red light)
    • jmIgnoreFoeProb and jmIgnoreFoeSpeed (ignore foes that are approaching the junction below the given speed with the given probability)
    • jmTimegapMinor < 1 (safety time gap when passing an intersection without priority)
    • jmIgnoreJunctionFoeProb > 0: allows ignoring foes that are already on the junction (with given probability)
      • vehicular foes as long as they are not in the way
      • any pedestrian foes
    • junctionModel.ignoreIDs: ignores all foes with the given ids (set via generic parameters)
    • junctionModel.ignoreTypes: ignores all foes with the given types (set via generic parameters)
  • yellow phases which are too short in relation to the vehicle speed (giving insufficient time to come to a stop). By default this causes strong braking (Warning: emergency braking) potentially followed by rear-end collisions
  • Green phases that allow conflicting streams to drive at the same time. Collision beyond the intersection due to this are always detected but collisions on the intersection are only registered when setting the option --collision.check-junctions.
  • If all Double connections at an edge are configured with pass="true"
  • If connections from the minor road of a priority junction are configured with pass="true"
  • Imperfect driver behavior

To simulate the effects of an accident see FAQ#How_to_simulate_an_accident.

Reaction Times#

In Reality, drivers take actions with some delay in regard to the evolving traffic situation. This reaction time is modelled in the simulation in various ways. Reaction time should not be confused with the tau parameter as this models the desired time headway. Keeping sufficient headway is necessary for safe driving with delayed reaction. However, this headway can be set independently of the actual reaction time.

Simulation Step Length#

All simulation vehicles select their speed simultaneously with regard to the traffic state from the previous simulation step. This means, there is at least a reaction time of the simulation step-length.

In contrast, lane-changing within an edge happens sequentially (in the upstream direction) within each step . Consequently, two vehicles that are driving almost parallel on a 3-lane road may both change their lanes in parallel within a single simulation step because the vehicle further behind already takes into account the lateral movement of the vehicle ahead.

Action Step Length#

By default, vehicles recompute their speed and lane-changing decisions every simulation step. When running with sub-second time resolution (i.e. --step-length 0.1), this also gives very low effective reaction times.

To decouple the decision interval from the simulation step length, the vehicle action-step-length can be defined as the duration between subsequent vehicle decisions. Decision-making starts directly after insertion which means vehicles inserted at different times may take decisions during different simulation steps. During simulation steps without decision-making, vehicle positions are updated according to the previously computed acceleration. Lateral dynamics (when using the sublane model) use a dynamic acceleration curve that to complete the current lateral driving manoeuvre.

The action step length can be defined in any of the following ways:

  • setting the option default.action-step-length <FLOAT> which sets the default action step length for all vehicles
  • setting the <vType> attribute actionStepLength
  • calling traci.vehicle.setActionStepLength

    Note

    the action step length must be a multiple of the simulation step length.

Safety-related parameters#

Simulation of unsafe behavior is a developing subject. Expect model additions and extensions. The following vehicle type parameters are safety related:

car-following model#

  • speedFactor: higher value causes speeding (driving above the speed limit)
  • accel, decel: higher values lead to abrupt speed changes
  • tau: lower values lead to reduced gaps to the leader vehicle. When setting tau lower than the simulation --step-length parameter (default 1s), resp. the vehicle's action step length (--default.action-step-length or vehicle parameter actionStepLength), collisions may occur
  • emergencyDecel: this is the maximum deceleration a vehicle can take to avoid an accident. By default this takes the same value as decel
  • apparentDecel: this is the deceleration value that will be assumed by other vehicles of the configured vehicle. By default this takes the same value as decel
  • driver imperfection modelled with the driverstate-device

Lane-Changing Model#

  • lcCooperative: lower values cause less cooperation during lane changing
  • lcSpeedGain: higher values cause more overtaking for speed
  • lcKeepRight: lower values cause less driving on the right
  • lcPushy: setting this to values between 0 and 1 causes aggressive lateral encroachment (only when using the Sublane Model)
  • lcImpatience: Repeated failure to change lanes causes lower lateral gaps to be accepted when using the sublane model
  • lcAssertive: setting this values above 1 cause acceptance of smaller longitudinal gaps in proportion to driver impatience(only when using the Sublane Model)
  • lcSigma: models random lateral variations (lane keeping imperfection)

Junction Model#

There are further parameters which affect safety-related junction behavior. For a description see Definition_of_Vehicles,_Vehicle_Types,_and_Routes#Junction_Model_Parameters.

  • impatience: higher values lead to acceptance of smaller time gaps at intersections. These are still safe in the sense of collision avoidance.
  • jmCrossingGap: lower values lead to more aggressive driving at pedestrian intersections (safe)
  • jmDriveAfterRedTime: Violate red lights some time after the switch (safety decreases with growing values)
  • jmDriveRedSpeed: Speed when violating red lights (higher values decrease safety)
  • jmIgnoreFoeProb: Probability to ignore foes with right-of-way (higher values decrease safety)
  • jmIgnoreFoeSpeed: Foe speed threshold which prevents fast vehicles from being ignored (Higher values decrease safety)
  • jmSigmaMinor: When planning to pass a minor link, optimum acceleration is assumed. (higher values cause reduced acceleration and thus decrease time gaps and safety)
  • jmTimegapMinor: Minimum time gap when passing a minor link ahead of a prioritized vehicle (lower values decrease safety)

Safety-Related Outputs#

  • Surrogate Safety Measures (SSM) (headway, brake rates, time to collision etc.)
  • The Instant Induction Loop output records the gap attribute which gives the time gap between successive vehicles with sub-second precision.
  • The Lanechange output generates a record of all lane-changes that take place in the simulation. It includes the longitudinal gap to the leader and follower vehicle on the target lane (both the actual gap and the one that would have been required to ensure stringent deceleration bounds under all circumstances). When using the sublane model, the attribute latGap records the lateral gap to the neighboring vehicle in the direction of the change if such a vehicle exists.
  • The collision output holds an xml record of each vehicle/vehicle and vehicle/person collision.