Splitting Gaussians in Mixture Models

Abstract

Gaussian mixture models have been extensively used and enhanced in the surveillance domain because of their ability to adaptively describe multimodal distributions in real-time with low memory requirements. Nevertheless, they still often suffer from the problem of converging to poor solutions if the main mode stretches and thus over-dominates weaker distributions. Based on the results of the Split and Merge EM algorithm, in this paper we propose a solution to this problem. Therefore, we define an appropriate splitting operation and the corresponding criterion for the selection of candidate modes, for the case of background subtraction. The proposed method achieves better background models than state-of-the-art approaches and is low demanding in terms of processing time and memory requirements, therefore making it especially appealing in the surveillance domain.

Foreground Segmentation with Tree-Structured Sparse RPCA

“…It proves itself to have excellent performance in dealing with heavy noise, thanks to the approximated RPCA model where the residual error (noise) is discarded into a third matrix in the decomposition. In addition, estimation of background motion induced by a jittering [19] .7173 (19) ------SPGFL [20] .9469 (3) -.8519 (5) .6988 (7) -.8156 (5) -SGMM [36] .8594 (18) .7251 (13) .6380 (18) .5397 (16) .7944 (14) .6481 (18) .7008 (17) ViBe+ [37] .8715 (17) .7538 (10) .7197 (11) .5093 (18) .8153 (9) .6646 (17) .7224 (16) SC-SOBS [38] .9333 (7) .7051 (16) .6686 (17) .5918 (12) .7786 (17) .6923 (16) .7283 (15) PCP+Alignment [35] .9109 (16) .7218 (15) .6941 (14) .5371 (17) .7885 (16) .7192 (12) .7286 (14) PSP-MRF [39] .9289 (8) .7502 (11) .6960 (13) .5645 (14) .7907 (15) .6932 (15) .7372 (13) PBAS [40] .9242 (13) .7220 (14) .6829 (15) .5745 (13) .8597 (6) .7556 …”

A Novel Video Dataset for Change Detection Benchmarking

“…Haines et al [33] also propose an automatic mode selection method, but with a Dirichlet process. A splitting GMM that relies on a new initialization procedure and a mode splitting rule was proposed in [34], [35] to avoid over-dominating modes and resolve problems due to newly static objects and moved away background objects while a multi-resolution block-based version was introduced in [36]. The GMM approach can also be expanded to include the generalized Gaussian model [37].…”

Background Subtraction on Embedded Hardware

“…Haines et al [58] also propose an automatic mode selection method, but with a Dirichlet process. A splitting GMM that relies on a new initialization procedure and a mode splitting rule was proposed in [46,48] to avoid over-dominating modes and resolve problems due to newly static objects and moved away background objects while a multi-resolution block-based version was introduced in [146]. The GMM approach can also be expanded to include the generalized Gaussian model [4].…”

“…This includes the well-known methods by Stauffer and Grimson [156], a self-adapting GMM by KaewTraKulPong [74], the improved GMM method by Zivkovic and Heijden [202], the multiresolution block-based GMM (RECTGAUSS-Tex) by Dora et al [146], GMM method with a Dirichlet process (DPGMM) that automatically estimated the number of Gaussian modes [58] and the SGMM and SGMM-SOD methods by Evangelio et al [48,46] which rely on a new initialization procedure and novel mode splitting rule. We also included a recursive per-pixel Bayesian approach by Porikli and Tuzel [138] which shows good robustness to shadows according to [54].…”