A WiSARD–based approach to CDnet Massimo De Gregorio and Maurizio Giordano Istituto di Cibernetica “Eduardo Caianiello” Consiglio Nazionale delle Ricerche Pozzuoli (NA) - Italy Email: {massimo.degregorio, maurizio.giordano}@cnr.it Abstract—In this paper, we present a WiSARD-based system (CwisarD) facing the problem of change detection (CD) in multiple images of the same scene taken at different time, and, in particular, motion in videos of the same view taken by a static camera. Although the proposed weightless neural approach is very simple and straightforward, it provides very good results in challenging with others approaches on the ChangeDetection.net benchmark dataset (CDnet). I. I NTRODUCTION Change Detection (CD) is the problem of detecting regions of change in multiple images of the same scene taken at different time. CD applied to videos taken from the same view by a static camera is an interesting problem in a large number of applications, such as video surveillance [1], remote sensing [2], medical diagnosis and treatment [3], just to mention a few. Change Detection problem is often referred in literature as Background Subtraction (BS) problem 1 , that is the task of distinguishing foreground objects from background areas in a video sequence. Several techniques have been proposed to CD problem and evaluation and comparison surveys of existing techniques can be found in literature [4], [5], [6], [7]. CD methods have to cope with different challenges and several problems arising from capturing scene frames from a static camera. A comprehensive list of CD challenges has been described in detail by Bouwmans et al. [5]. In our work 2 , we restrict our CwisarD method evaluation to the challenges addressed by the ChangeDetection.net website competition and its dataset repository [8]. Most of CD approaches can be classified into the follow- ing main categories: pixel–based [9], region–based [10] and object–based [11]. The first one mainly based on individually pixel changes. The second one based on the analysis of the single pixel with its neighborhood. The latter based on splitting the image into regions that are likely to belong to the same object. Another classification of CD methods has been proposed by Bouwmans in his survey [5], among these methods there are those based on Neural Network Modeling, such as, for instance: self organizing neural network [12], general regres- sion neural network [13], adaptive resonance theory neural network [14]. Even if the CD problem has been approached by different neural architectures, the totality of them is based on a 1 Also called Foreground Detection problem. 2 The research leading to these results has received funding from the EU FP7-ICT-2012-8 under the MIDAS Project - Grant Agreement no. 318786). weighted neuron model. On the contrary, in this paper we face the change detection problem with a weightless neural system (RAM–neuron model) that we called CwisarD. The approach is very straightforward, the image pre– and post–processing are very simple and the system bases its elaboration on single pixel information to accomplish the CD task. The paper is so organized: in Section II the CwisarD weight- less neural approach to CD problem is represented; Section III describes performance metrics and experimental settings we used to evaluate CwisarD performance; Section IV reports and discusses the experimental results of CwisarD detection capabilities when running on the ChangeDetection.net bench- mark dataset; Section V describes a parallel implementation of CwisarD and evaluates its time performances; finally Section VI summarizes concluding remarks and future perspectives of CwisarD. II. THE WI SARD APPROACH TO CD WiSARD has been the first commercial neural machine and was introduced by Aleksander et al in the early 80’s [15]. WiSARD is composed by a given number of discriminators, each one representing a different class. Each discriminator is built from Xn-tuple RAM nodes (one-bit words), all initially set to “0”. These RAM nodes are commonly called neurons. During the training phase, a X × n bits binary input pattern is given to the corresponding discriminator so that all addressed RAM memory locations are set to “1”. In the classification phase, the sum of all the memory contents addressed by a given input pattern represents each discriminator response. Such input pattern is associated to the discriminator class whose response is the highest (see [16]). In order to feed the discriminators with the right input, CwisarD creates one discriminator for each pixel of the image (see Fig. 1). The RGB color of the pixel is represented by a binary (black and white) image, where the colums represent Fig. 1: CwisarD discriminator input