Background Subtraction Based on Modified Pulse Coupled Neural Network in Compressive Domain

Background subtraction is the most popular technique for foreground detection to automatically separate foreground objects from background in video sequences. Much research effort has been directed into exploiting different features for background modeling to improve robustness of background models. However, most features used for background modeling are local features (such as intensities, colors and textures) which may have limitations when dealing with environments with complex and diverse changes in backgrounds, since changes may not emerge only within local regions. In this paper, inspired by mature human visual system (HVS), we propose a background subtraction method based on modified pulse coupled neural network (PCNN) in compressive domain, extracting global information for background modeling by making use of global coupling characteristics of the PCNN. Enlightened by theory of compressive sensing, we first employ compressive projection matrices to reduce data dimensionality. Then, we modify the PCNN to receive multi-channel color information as external stimuli and adjust linking strength and linking range for each neuron to enhance the ability of the network in global coupling. The modified PCNN is utilized to generate output responses containing global information of scenes for background modeling. Experimental results have demonstrated that our method has excellent performance and outperforms several advanced techniques in terms of visual and quantitative evaluations.