Instance-vote-based motion detection using spatially extended hybrid feature space

Abstract

Motion recognition, a trivial step employed in several video-based applications, is still a challenging task in real-world complex scenarios containing dynamic noise, varying backgrounds, shadows, improper illuminations, camouflages, etc. Numerous pixel-based change detection techniques employing varied combinations of different feature spaces have been proposed to efficiently overcome many real-world challenges. But ideally, handling all the possible real-world challenges simultaneously is yet to be achieved. Hence, this paper proposes a memory-efficient unique combination of multi-colour feature space with a light-weight intensity-based texture descriptor. The proposed spatially enlarged extended centre-symmetric local binary pattern is combined with YCbCr and RGB colour features for robust pixel representation. The proposed feature space is fed to an extended instance-vote technique for pixel classification. The random and time-subsampled update is employed conditionally for model update, followed by a feedback network that continuously optimizes the local threshold and learning rate parameters of the proposed model. The proposed feature space and model have been evaluated on whole 2014 Change Detection dataset, the largest known dataset. The outperforming performance and memory analysis strengthens its acceptability for real-time applications.