We propose an algorithm for detecting objects that present potential hazards to drivers. The input data is a combination of local information derived from optical flows and global information obtained from the host vehicle's status. We use artificial neural networks to infer the degree of danger posed by moving objects captured in dynamic images taken with a vehicle-mounted camera. Our approach allows for adapting the algorithm flexibly to numerous drivers of dissimilar characteristics.

To test our algorithm, we conducted experiments with miniature vehicles in virtual environments and with real vehicles in real driving situations, using video images of moving objects. The results verify that the algorithm can infer hazardous situations in a manner similar to the judgments made by human drivers. Our proposed algorithm thus offers a foundation for constructing a practical driving-assistance system, and an automobile manufacturer in Japan is studying possible applications for the algorithm.