Foreground Object Detection Based on Multi-model Background Maintenance

Abstract

This paper addresses the problem of background maintenance for foreground object detection. A Multimodel Background Maintenance (MBM) framework that contains two principal features is proposed. Under this framework, a pure time-varying background image is maintained and learned using the statistical information of the multi-model Gaussian distribution with principle features. The principal features consist of static and dynamic pixels to represent the characteristic of background. Experiments are conducted on ten image sequences containing targets of interest in a variety of environments. Quantitative evaluation and comparison with the existing method show that the proposed method provides much improved results. Ninth IEEE International Symposium on Multimedia 2007 -Workshops 0-7695-3084-2/07 $25.00

Design of Real-Time Self-Frame-Rate-Control Foreground Detection for Multiple Camera Surveillance System

“…The quantitative evaluation result is given in Table 3. The performance of proposed system much approaches to the performance of [5] after post-processing. As comparing with a more sophisticated algorithm [10], the results of proposed algorithm are generally lower than the results of [10] before post-processing.…”

“…In the following experiments, these two parameter settings are applied to all indoor type of sequences and all outdoor type of sequences respectively. Detection performance evaluations: the foreground detection results are compared with a FD based algorithm [19], the GMM algorithm [3], the MBM [5], and a LBP based algorithm [10]. To show the effect on the updating by LUTGaussian PDF, TDCP technique, and frame rate control scheme, both the foreground detection results without postprocessing and with post-processing are obtained.…”

“…The foundation of the proposed foreground detection method is a multiple background maintenance method [5] and a statistic prediction method, namely TDCP. The multiple background maintenance method models the recent history of observation at a given position as a mixture of K Gaussian clusters, and each Gaussian θ associates with weight ω, mean μ and variance σ 2 .…”

“…The soft-partition update strategy [6] is adopted instead of winner-take-all principal as that in [3]- [5]. The softpartition updates all Gaussians by an amount proportional to their posterior probabilities.…”

Monocular vision-based depth map extraction method for 2D to 3D video conversion

“…T is a threshold that compares with the difference between the current pixel and ith Gaussians mean values. T is also estimated depending on the specified probability PR [24]. Figure 4c shows the result of moving object detection.…”