Evaluation of an Adaptive Composite Gaussian Model in Video Surveillance

Abstract

Video surveillance systems seek to automatically identify events of interest in a variety of situations. Extracting a moving object from background is the most important step of the whole system. There are many approaches to track moving objects in a video surveillance system. These can be classified into three main groups: feature-based tracking, background subtraction, and optical flow techniques. Background subtraction is a region-based approach where the objective is to identify parts of the image plane that are significantly different to the background. In order to avoid the most common problems introduced by gradual illumination changes, waving trees, shadows, etc., the background scene requires a composite model. A mixture of Gaussian distributions is most popular. In this paper, we classify and discuss several recently proposed composite models. We have chosen one of these for implementation and evaluate its performance. We also analyzed its benefits and drawbacks, and designed an improved version of this model based on our experimental evaluation. One stationary camera has been used.Keywords: video sequence analysis, surveillance systems, Gaussian mixture models. You are granted permission for the non-commercial reproduction, distribution, display, and performance of this technical report in any format, BUT this permission is only for a period of 45 (forty-five) days from the most recent time that you verified that this technical report is still available from the CITR Tamaki web site under terms that include this permission. All other rights are reserved by the author(s). Evaluation of an Adaptive Composite Gaussian Model in Video Surveillance

On the Mahalanobis Distance Classification Criterion for Multidimensional Normal Distributions

“…According to the sixth column, two categories may be distinguished: works that provide an incorrect confidence level [9][10][11] [33], and works that are incomplete (marked with '-') because they do not specify such confidence values. According to the third column, many works [20][21] [22][23] [24] are ambiguous about the classification criterion used. In most cases, this is due to the ambiguity inherited from [13].…”

“…If the ambiguity is resolved, two interpretations are possible: some treatises [14][16] [33] opt for hyper-rectangular regions, while others [17][19] choose ellipsoidal ones. In other references it is not possible to infer the criteria used, and the choice between both criteria causes confidence variations, e.g., a 6% variation for the n = 3, d = 2.5 case [13] [20][22] [24]. The last two columns of Table I are given taking into account both possible disambiguations.…”

“…Other approaches [17] [18] [19] use the Mahalanobis distance to the mean of the multidimensional Gaussian distribution to measure the goodness of fit between the samples and the statistical model, resulting in ellipsoidal confidence regions. Finally, a third group of treatises [20][21] [22][23] [24] mimic the description in [13] and therefore do not disambiguate the shape of the decision region. Similarly to [13], most of these works set the classification threshold to 2.5 regardless of the dimensionality of the Gaussian model and the shape of the confidence region considered.…”

Robust background subtraction and maintenance

“…This paper uses a MOGS for modelling static scenes. We discussed in [11] that a MOGS does have some advantages, such as adaptivity, time-efficiency and robustness. But there are further problems due to shadows, camouflage, or very large and very slow moving objects.…”

Improved Parameters Updating Algorithm for the Detection of Moving Objects

“…Hybrid models such as GMM and K-means [3], GMM and fuzzy logic [1], Markov Random Fields [22],GMM and adaptive background [9,25], have been proposed to overcome GMM drawbacks. Other works have focused on improving the learning speed [15,28] through an adaptive learning rate [29], Better settings White and Shah [32] and the execution time [17] by using real parallel operations on multi-processor machines. Other systems use two backgrounds [4] to solve the problem of change in brightness between day and night or use Multi-level approaches [5][6][7].…”