Robust event detection by radial reach filter (RRF)

We propose a novel statistical measure for robust event detection, called ´Radial Reach filter´ (RRF). The capability of detecting new objects (events) from a time-series image is an important problem of vision systems. The usual method of detecting new objects is simple background subtraction, that is to subtract current image from a background image. However, simple background subtraction is susceptible to illumination change such as shadows. Moreover, when the brightness difference between events and a background is small, it cannot detect the difference. In order to solve such problems, we propose the RRF which evaluates a local texture and realizes robust event detection. Experiments using real images show the effectiveness of the proposed methods. Furthermore, an experiment using an all-directional image from a stereo omni-directional system (SOS) shows the possibility of application to an environment-monitoring system.