Multi-scale multi-feature codebook-based background subtraction

Abstract

This paper presents a novel real-time multi-feature multi-scale codebook-based background subtraction algorithm, targeted for challenging surveillance environments. Our contribution is three-fold. First, we present an extension of the Codebook background model [4] that combines multiple features, such as intensity, colour and texture, in a principled way, simultaneously taking into account both the feature's confidence and its similarity score. Second, a new local texture pattern descriptor is proposed, entitled Local Ratio Pattern, generalizing previously successful local pattern methods [9]. Third, a generic multi-scale confidence fusion scheme is provided, in order to aggregate individual results at different scales. A thorough evaluation is performed on the challenging I2R dataset [8]. In addition, a comparison is carried out with other competing methods, leading to state-of-the-art performance.