Detectability Prediction for Increased Scene Awareness

A driver assistance system realizes that the driver is distracted and that a potentially hazardous situation is emerging. In this scenario the driver needs to make an optimal decision as fast as possible. His attention needs to be directed to the location that enhances the perception of all action relevant entities. But where is that optimal spot? Pedestrian detectability is a measure of the probability that a driver perceives pedestrians in static and dynamic scenes. Leveraging this information allows a driver assistance system to direct the attention of the driver to the spot that maximizes the probability that all pedestrians are seen. In this paper we present a thorough description and a strong theoretical foundation of this concept. We use annotated datasets recorded in urban environments, and acquire the detectabilities of pedestrians via psychophysical experiments. Based on these measured detectabilities, we train a machine learning algorithm to predict detectability from an optimized set of image features. We furthermore exploit this mapping to obtain the optimal focus of attention, thus demonstrating the potential benefit of our method in a realistic driver assistance context.