PURPOSE: To prevent the collision by predicting a car collision critical state and taking a prescribed safety countermeasures.

CONSTITUTION: The picture of a scenery within a visible range from a self car picked up just before the collision is inputted to an input layer 1 of a neural network having a hidden layer 2 as the group of learning data and the characteristic is extracted and learnt. Real picture data collected by an on-vehicle image pickup means in real time at the time of traveling are inputted to a recognition section having the collision prediction circuit equipped with a learnt neural network as a data set in succession. In this case, when succeeding pictures are judged to be those of collision, an collision flag is outputted to the output value of an output layer 3. By accepting this, the operation start of the vehicle traveling safety holding means is instructed immediately.