Reducing false alarm rates in surveillance imaging using significance testing

We describe a technique for detecting changes in an imaged scene that is robust to noise and to natural random motions of fixed objects (e.g. trees) in the scene. For a static camera much of a scene will remain constant over time periods that are very long compared to the time between images. This is true to varying degrees for all regions of an image. Changes in the structure of an image may be of two types; those due to “natural motion” (typically trees, bushes, pedestrians, vehicles) which are not significant, and those due to significant changes (which correspond to alarm conditions, e.g. intruders). Significant changes are those that stand out from the background of normal activity in the image. In our application we compute the deviation of a pixel from its long-term mean value, this value is squared and then normalised by the variance of that pixel. The sensitivity is reduced, pixel by pixel, across the image if the underlying variability of the image is high at that pixel. The price for this is reduced sensitivity to an “intruder” in image regions of high variance, the benefit is reduced levels of false alarms while retaining high sensitivity in image regions of low variance