How does a driverless car predict possible collisions in a complex city traffic?

The control software first builds a model of the surrounding. This means it uses the pictures from the cameras and the data from radar and lidar to derive the location, the velocity and the acceleration of the own car and all the other cars, bicycles, pedestrians, animals, socceer balls etc. on the road. This model is then used to predict the movement of all traffic participants on the road.

Since the acceleration and deceleration is limited by physical boundaries (i.e. how slippery is the raod), the possible outcomes of the next seconds can be predicted taking different reactions of the participants into account. The bad situation is, if the outcome of the prediction is a crash independent of possible reactions of the participants. Then the control software tries to limit the damage. The strategy for the limitation is a point of big debate. Who will be favoured by the strategy? The driver or the other drivers or an average solution?

There are also simpler versions of control software available, which are strongly based on neural networks for the derivation of the control commands (stearing and acceleration/deceleration). These simpler versions are currently not built to predict accidents. They just try to avoid contact with other traffic participants.