The SOBS algorithm: What are the limits?

Abstract

The Self-Organizing Background Subtraction (SOBS) algorithm implements an approach to moving object detection based on the neural background model automatically generated by a self-organizing method, without prior knowledge about the involved patterns. Such adaptive model can handle scenes containing moving backgrounds, gradual illumination variations and camouflage, can include into the background model shadows cast by moving objects, and achieves robust detection for different types of videos taken with stationary cameras. Moreover, the introduction of spatial coherence into the background update procedure leads to the so-called SC-SOBS algorithm, that provides further robustness against false detections. The paper includes extensive experimental results achieved by the SOBS and the SC-SOBS algorithms on the dataset made available for the Change Detection Challenge at the IEEE CVPR2012

Changedetection.net: A new change detection benchmark dataset

“…It should come as no surprise that the three simplest methods based on plain background subtraction [4,11] are at the bottom of the table, whereas the four most recent methods [22,7,10,14] are at the top. The methods ranked number 1 [10] and number 3 [7] are closely related.…”

“…As shown in the paper, this approach reduces both the FNR and the FPR of any method it is used on. As for the fourth ranked method SC-SOBS [14], its approach is orthogonal to traditional motion detection methods in its use of a self-organizing neural network. Such an approach gives remarkable results on baseline and intermittent object motion videos.…”

“…Two fairly old, but frequently-cited, methods: KDE-based estimation by Elgammal et al [8] and GMM by Stauffer and Grimson [24], as well as 5 improved versions of these methods: self-adapting GMM by KaewTraKulPong [12], improved GMM by Zivkovic and van der Heijden [28], block-based GMM by Dora (RECTGAUSS-Tex) et al [20], multi-level KDE by Nonaka et al [17], and spatio-temporal KDE by Yoshinaga et al [1] were also tested. We also include results for a machine learning method based on neural maps (SOBS and SC-SOBS) by Maddalena et al [13,14], a post-processing method based on probabilistic super-pixels (PSP-MRF) [22], a fairly complex commercial method that does pixel-level detection using the Chebyshev inequality and peripheral and recurrent motion detectors by Morde et. al.…”

Background Subtraction on Embedded Hardware

“…Results for data-driven methods and machine learning methods are also reported. That is Hofmann's stochastic and self-adaptive method (PBAS) [64], a simple K-nearest neighbor method [202] and neural maps methods (SOBS and SC-SOBS) by Maddalena et al [104,106] and a neural network method with a region-based Markovian post-processing methods (PSP-MRF) by Schick et al [152]. We also have results for two commercial products.…”

“…This method do not need pre-labeled training data, but also allows for online updating of the SVM parameters. Maddalena and Petrosino [104,106] model the background of a video with the weights of a neural network. A very similar approach but with a post-processing MRF stage has been proposed by Schick et al [152].…”