A pixel-level, intensity-based nonlinear autoregressive classifier (NARX) with chromatic exogenous input for efficient image background subtraction

Background subtraction is a well-known technique in computer vision to extract foreground objects from background reference frames. In real-time video processing applications such as surveillance, behavioral profiling and intelligent transport systems, the domain presents a number of challenges. Video frames used to train such models contain a range of dynamic background activities such as waving trees, moving cloud cover or abrupt intensity variations that make the foreground detection a challenging task. Dynamic neural networks are known for their capability to predict time-series-based nonlinear models via previous feature data. The proposed scenario models each pixel´s intensity/color-alternating behavior based on its previous activity patterns. Any significant or unusual variation in the underlying intensity or color value therefore is modeled as a foreground activity. Based on this concept, this paper presents a non-linear autoregressive neural (BG-NARX) classifier with the pixels´ chromatic values as the exogenous vectors to improve background detection accuracy. The proposed model was evaluated against three benchmarking video datasets and reported promising detection accuracies ranging from 67-94% for pedestrians and vehicles against highly variable backgrounds with low false positives and negatives.