A recognition system that uses saccades to detect cars from real-time video streams

In this work we present a system for detection of objects from video streams based on properties of human vision such as saccadic eye movements and selective attention. An object, in this application a car, is represented as a collection of features (horizontal and vertical edges) arranged at specific spatial locations with respect to the position of the fixation point. During the recognition process, the system efficiently searches the space of possible segmentations by investigating the local regions of the image in a way similar to human eye movements. In contrast to motion-based models for vehicle detection, our approach does not rely on motion information, and the system can detect both still and moving cars in real-time.