CwisarDH+: background detection in RGBD videos by learning of weightless neural networks

This work is a continuation of our past experiences in investigating learning and classification capabilities of weightless neural networks (WNNs) for background modeling and object motion detection in videos. In the current work, we adapted and modified a previous method, called CwisarDH, to the domain of RGBD videos. In the proposed approach two main strategies were adopted. First, by decoupling the RGB color information from the pixel depth information, the two video streams are synchronously but separately (under different neural configurations) modeled by WNNs at each pixel. Depending on the video temporal ROI, a preliminary set of frames is used solely for network training. In the detection phase, classification is interleaved with re-traning on current colors whenever pixels are detected as belonging to the background. Secondly, the independent outputs of the two video processing are combined by an OR operator and post-processed by erosion/dilation filters. With this simple approach, we obtained an efficient background modeling in RGBD videos, as we were confident of this considering the good results gathered by CwisarDH in the ChangeDetection.net 2014 challenge.