Toward a computational approach for collision avoidance with real-world scenes

摘要

In the central nervous systems of animals like pigeons and locusts, neurons were identified which signal objects approaching the animal on a direct collision course. In order to timely initiate escape behavior, these neurons must recognize a possible approach (or at least differentiate it from similar but non-threatening situations), and estimate the time-to-collision (ttc). Unraveling the neural circuitry for collision avoidance, and identifying the underlying computational principles, should thus be promising for building vision-based neuromorphic architectures, which in the near future could find applications in cars or planes. Unfortunately, a corresponding computational architecture which is able to handle real-situations (e.g. moving backgrounds, different lighting conditions) is still not available (successful collision avoidance of a robot was demonstrated only for a closed environment). Here we present two computational models for signalling impending collision. These models are parsimonious since they possess only the minimum number of computational units which are essential to reproduce corresponding biological data. Our models show robust performance in adverse situations, such as with approaching low-contrast objects, or with highly textured backgrounds. Furthermore, a condition is proposed under which the responses of our models match the so-called eta-function. We finally discuss which components need to be added to our model to convert it into a full-fledged real-world-environment collision detector.