Local Spatial Co-occurrence for Background Subtraction via Adaptive Binned Kernel Estimation

Abstract

We present a nonparametric background subtraction method that uses the local spatial co-occurrence correlations between neighboring pixels to robustly and efficiently detect moving objects in dynamic scenes. We first represent each pixel as a joint feature vector consisting of its spatial coordinates and appearance properties (e.g., intensities, color, edges, or gradients). This joint feature vector naturally fuses spatial and appearance features to simultaneously consider meaningful correlation between neighboring pixels and pixels’ appearance changes, which are very important for dynamic background modeling. Then, each pixel’s background model is modeled via an adaptive binned kernel estimation, which is updated by the neighboring pixels’ feature vectors in a local rectangle region around the pixel. The adaptive binned kernel estimation is adopted due to it is computationally inexpensive and does not need any assumptions about the underlying distributions. Qualitative and quantitative experimental results on challenging video sequences demonstrate the robustness of the proposed method.