Evaluation of a Self-learning Event Detector

Abstract

Recently, a novel solution philosophy for complex pattern recognition problems has been proposed (R. Mattone et al. in [1]), which is suitable to be applied in the video-surveillance context. It is based on the use of low-level image features and simple, local statistical models for the recognition of complex patterns. For example, the amount of motion in a scene, represented by the number of pixels that changed their intensity over a short time, can be used to recognise “unusual” behaviour of persons like panic or overcrowding. Obviously, this concept calls for some kind of model describing the “normal” situation (e.g., concerning motion) in the field of view. This has been realised by a suitable decomposition of the scene into regions where local models of “normality” can be built. These novel concepts have been implemented in a commercial application called SEDOR, developed at Ascom AR&T. We describe here recent systematic tests of SEDOR, performed on more than 20 hours of video-tapes to demonstrate its performances. We will illustrate that the results can be improved, and point out a possible solution.