Boost sparse coding based abnormal event detection via explicitly applying temporal continuity constraint

Abstract

With various novel methods having been proposed, recent years witnessed great progress in abnormal event detection. Broadly speaking, most existing methods can be divided into two categories: global feature representation based ones and local feature representation based ones, though the specific feature model and scale differ a lot. These two types of methods have reverse pros and cons: global feature representation methods can better guarantee spatial-temporal continuity of abnormal events but lack the ability to accurately model features of the basic event elements, while local feature methods are just the opposite. That makes their results complement each other. In this paper, we propose to explicitly apply temporal continuity constraint on sparse coding based local feature representation method, not just enlarging the scale of local feature representation. Experiments demonstrate that our method can usually achieve more stable and smooth results, thus more high detection accuracy. In some cases, the performance gain can be enormous.