Abstract
Keywords: autonomous driving, collision probability, collision avoidance, decisionmaking, spacetime
This paper describes a new stochastic decision-making method for use in an autonomous driving system that minimizes the probability of a collision.
One of the most important tasks for an autonomous driving system is collision avoidance. This paper defines a collision probability for each trajectory of an autonomous vehicle in consideration of all possible trajectories of other objects. This probability is used to find the best trajectory to help avoid collision for the autonomous vehicle.
The collision probability is computed in 3-dimensional spacetime, where the car's (x, y) position at time t is the point (t, x, y) in 3D spacetime. Similarly, the movement of the car during t=0 to t=T is represented as a trajectory in 3D spacetime.
When the movement is only known statistically, it is represented as a set of possible future trajectories with their probabilities. The probability of each trajectory is computed as a product of the probabilities of control actions of each time step, such as accelerations and angular velocities. If there is no knowledge about operations' probabilities in advance, the uniform probability is assumed because it represents the most unpredictable case.
To compute the collision probability between an autonomous vehicle with a single trajectory and another car with multiple possible trajectories, we sum the probabilities of the other car's trajectories that intersect with the autonomous vehicle's trajectory. If there are multiple cars, they are each represented as sets of trajectories, and the probability that an autonomous vehicle will collide with at least one car is used as the collision probability for each autonomous vehicle's trajectory.
The safest trajectory for an autonomous vehicle is the trajectory that minimizes the collision probability defined above.
This approach has been tested in a scenario where the autonomous vehicle is merging onto an expressway. Our tests show that the autonomous vehicle will choose appropriate actions as follows:
1) accelerate to match cruising lane velocities (100km/h),
2) merge into the cruising lane, keeping appropriate distance from other cars,
3) merge into the passing lane and go forward while keeping appropriate distance from other cars.
These results indicate that the collision probability measure introduced here is well suited for this application, and minimizing this probability is helpful for collision avoidance. This measure should be considered in the design of autonomous driving systems.