Adaptive background modeling of complex scenarios based on pixel level learning modeled with a retinotopic self-organizing map and radial basis mapping

摘要

Background modeling in video sequences is a prominent topic which generates very relevant works regarding models, algorithms, and databases. Its importance is related to real world applications like, video segmentation, surveillance, Internet of things (IoT), privacy and video compression. This paper proposes an adaptive background modeling method termed Radial Basis RESOM (RB-SOM) in order to contribute in this field. RB-SOM is able to deal with video sequences involving complex scenarios. The model assigns to each pixel an adaptive learning rate, which is determined by two schemes: a mapping radial basis function, and a retinotopic self-organizing neural network (RESOM). The radial function defines information to activate the learning of pixels representing the background and inhibit the learning of pixels that correspond to dynamic objects, even when they stay still for a long time. RESOM identifies the gradual illumination changes and improves the separation of background and foreground. The model also performs an entropy analysis and correlation to detect abrupt scenario changes such as sudden illumination changes. Results of the work indicate that the performance of the proposed model is similar to state of the art background modeling algorithms like BEWiS and LaBGen, and compared to motion detection models like Subsense and PAWCS. Then, RB-SOM presents the best processing times, and stabilization of the background when the video sequence presents sudden scenario changes. Besides, RB-SOM has competitive results in videos with dynamic background, jittering, illumination changes, and dynamic objects that remain stationary.