#### DMCA

## PROBABILISTIC METHODS FOR ADAPTIVE BACKGROUND SUBTRACTION (2008)

Citations: | 2 - 0 self |

### Citations

8745 |
Controlling the false discovery rate: a practical and powerful approach to multiple testing
- Benjanimi, Hochberg
- 1995
(Show Context)
Citation Context ...tive in space. Now, we shall present a complementary detection method which adapts in time. The method is a multiple comparisons procedure (MCP) that controls the socalled false discovery rate (FDR) (=-=Benjamini and Hochberg, 1995-=-). The false discovery rate is defined as the expected proportion of false alarms to total positive declarations. Recall from Section 3.2 the definition in terms of the random variables V and R is FDR... |

5126 | Stochastic relaxation, Gibbs distributions, and the Bayesian restoration of images
- Geman, Geman
- 1984
(Show Context)
Citation Context ...∣∣ eB} Pr {I = I[n] ∣∣ eF} ≡ PB(I[n])PF (I[n]) . Now we focus on the right-hand side of (5.3). Since E is a Markov random field, the a priori probabilities are Gibbsian of the following general form (=-=Geman and Geman, 1984-=-): Pr {E = e} = 1 Z exp ( −1 T ∑ c∈C V(c) ) (5.5) where Z and T are normalization and natural temperature constants respectively. While Z gets canceled in the prior ratio, T is a parameter to be set b... |

1482 | Pfinder: realtime tracking of the human body
- Wren, Azarbayejani, et al.
- 1997
(Show Context)
Citation Context ...foreground statistics more precisely, one may expect greater discrimination ability. In some specific applications, such as body part segmentation, foreground object region growing has been done. In (=-=Wren et al., 1997-=-) and in (Elgammal et al., 2002), foreground regions are morphologically grown according to specific shape and color models of the head, torso, and extremities. In both of these works however, the for... |

1248 | On the statistical analysis of dirty pictures,”
- Besag
- 1986
(Show Context)
Citation Context ... field represents our initial “best guess” of the true result. In this work, we have used a deterministic relaxation method to optimize the detection result known as iterated conditional modes (ICM) (=-=Besag, 1986-=-). An alternative approach would be to use a stochastic relaxation technique, such as simulated annealing, which will converge to the true result with high likelihood (Geman and Geman, 1984). This is ... |

898 | Learning patterns of activity using real-time tracking.
- Stauffer, Grimson
- 2000
(Show Context)
Citation Context ...poses a slight problem, however, since it is not clear how to initialize the model or how to choose an appropriate number of mixture components to begin with. An on-line K-means approach is taken by (=-=Stauffer and Grimson, 2000-=-). In this approach, each incoming pixel is matched to a cluster if it is within 2.5 standard deviations from the cluster mean. The parameters for that cluster are then updated with the new observatio... |

867 |
Testing Statistical Hypotheses,
- Lehmann
- 1986
(Show Context)
Citation Context ...common definition of a p value from the Neyman-Pearson statistical testing viewpoint is the probability that a less likely outcome than the current observation could occur given 59 a null hypothesis (=-=Lehmann, 1986-=-). In our treatment, this general definition is followed faithfully. Interpreted mathematically, this translates to p = f(x) = ∫ R1(x) P0(y)dy R1(x) = {y ∣∣ P0(y) ≤ P0(x)}. (6.1) This definition is il... |

699 | Probabilistic visual learning for object representation.
- Moghaddam, Pentland
- 1997
(Show Context)
Citation Context ...onto this low dimensional subspace. The residual reconstruction error, or “difference from feature space” (DFFS), defined as ǫ2(I) = ||I − µb||2 − M∑ m=1 ΦTM (I − µb) is thresholded to detect motion (=-=Moghaddam and Pentland, 1997-=-). The authors of (Oliver 13 et al., 2000) cite that the eigenbackground approach has lighter computational load than an MoG type of description. However, they also rely on a morphological post proces... |

680 |
Pattern Recognition,
- Theodoridis, Koutroumbas
- 1999
(Show Context)
Citation Context ...ound subtraction on an image I[n] entails applying a label, either B or F , to each pixel by comparing the current image to some background model. The problem can be formulated as hypothesis testing (=-=Theodoridis and Koutroubas, 2006-=-). In 15 the common fashion, the decision is framed thusly: Pr {I[n] = I[n] ∣∣ B} Pr {I[n] = I[n] ∣∣ F} B≷F ηπFπB . (3.1) The probabilities on the left hand side denote the probability of observing th... |

545 | Non-parametric model for background subtraction,” in
- Elgammal, Harwood, et al.
- 2000
(Show Context)
Citation Context ...raction literature to-date, the methods described utilize some kind of pixel-based probability assignment. For this work, the primary method for modeling the background is via the method proposed in (=-=Elgammal et al., 2000-=-). This method was chosen because of it’s accuracy, generality, and simplicity. The main contributions presented in this thesis are within the realm of background detection and they apply generally to... |

538 | A bayesian computer vision system for modeling human interactions,”
- Oliver, Rosario, et al.
- 2000
(Show Context)
Citation Context ... Bhattacharyya distances, this method will be slow and require a large amount of memory. Another interesting approach to background subtraction is through the calculation of “eigenbackgrounds” as in (=-=Oliver et al., 2000-=-). Each of N training images (i.e., containing only background) are represented as vectors: if the images are natively (H×W ) in dimension, the vector representation is (HW × 1). The mean of each pixe... |

477 | Wallflower: Principles and Practice of Background Maintenance”, Computer Vision,
- Toyama, Krumm, et al.
- 1999
(Show Context)
Citation Context ...their final blob segmentations of pedestrians before higher level processing is performed. The “Wallflower” algorithm, designed for body part segmentation and gesture recognition for is presented in (=-=Toyama et al., 1999-=-). The initial phase of background subtraction in that work was based on a temporal Wiener filter at each pixel. Based on 50 previous observations, a linear prediction of the current observation is co... |

294 | L.S.: Background and foreground modeling using nonparametric kernel density estimation for visual surveillance.
- Elgammal, Duraiswami, et al.
- 2002
(Show Context)
Citation Context ...ther words, a probability density function is estimated at each location. As discussed in Chapter 2, there are a number of ways that this can be done. Following the example of (Elgammal et al., 2000; =-=Elgammal et al., 2002-=-), this work considers models of the non-parametric type. The non-parametric estimate is attractive because it converges to the true underlying PDF with an appropriately selected kernel function (Theo... |

256 | Image Segmentation in Video Sequences: A Probabilistic Approach", - Friedman, Russell - 1997 |

248 | Detecting moving objects, ghosts, and shadows in video streams,”
- Cucchiara, Grana, et al.
- 2003
(Show Context)
Citation Context ...arly has the advantage of being both very simple and memory light, since no buffer is required. The detection performance with this method, however, is poor compared to the sliding window median. In (=-=Cucchiara et al., 2003-=-), an object-based approach is taken whereby moving objects, so-called “ghosts”, and shadows thereof are detected based on a number of features of detected foreground regions (size, optical flow, inte... |

236 | Image change detection algorithms: A systematic survey.
- RADKE, ANDRA, et al.
- 2005
(Show Context)
Citation Context ...Background subtraction is a general term for a process which aims to segment moving foreground objects from a relatively stationary background. As it has been noted in (Migdal and Grimson, 2005) and (=-=Radke, 2005-=-), we point out that that there is an important distinction between the background modeling and background detection stages, which comprise the whole subtraction process. As illustrated in Figure 1.1,... |

129 |
Video processing and communications”,
- Wang, Ostermann, et al.
- 2001
(Show Context)
Citation Context ...vely. Pan, tilt, and roll designate rotation about the vertical, horizontal, and optical axes respectively. Two-dimensional image motion models corresponding to these camera motions are described in (=-=Wang et al., 2002-=-) and (Stiller and Konrad, 1999). The motion models are essentially mapping functions that describe the displacement of all points in the image resulting from camera motion. A camera undergoing transl... |

99 | Statistical model-based change detection in moving video.
- Aach, Kaup
- 1993
(Show Context)
Citation Context ... dependence on neighboring labels can be expressed through a Markov random field (MRF) model. Markov random field models have been used in image and video processing in a variety of applications. In (=-=Aach et al., 1993-=-) and (Aach and Kaup, 1995), MRF models are applied to change detection masks of image pairs: e.g., subsequent frames in video. In those treatments, the test statistics were based on pixel-wise image ... |

51 |
Automatic congestion detection system for underground platform,
- Lo, Velastin
- 2001
(Show Context)
Citation Context ...mean and parameterized by variance σ2j . This concept may be extended further to include other image features such as intensity (color) gradients and edges which may offer greater discrimination. In (=-=Lo and Velastin, 2001-=-), for example, the input images to the background subtraction algorithm are first fed through what the authors dub a “variance filter”. This feature extractor computes, for each pixel, the variance o... |

50 |
Bayesian algorithms for adaptive change detection in image sequences using Markov random fields.
- Aach, Kaup
- 1995
(Show Context)
Citation Context ...ng labels can be expressed through a Markov random field (MRF) model. Markov random field models have been used in image and video processing in a variety of applications. In (Aach et al., 1993) and (=-=Aach and Kaup, 1995-=-), MRF models are applied to change detection masks of image pairs: e.g., subsequent frames in video. In those treatments, the test statistics were based on pixel-wise image differences. In this thesi... |

47 | Estimating motion in image sequences: A tutorial on modeling and computation of 2D motion,
- Stiller, Konrad
- 1999
(Show Context)
Citation Context ...l designate rotation about the vertical, horizontal, and optical axes respectively. Two-dimensional image motion models corresponding to these camera motions are described in (Wang et al., 2002) and (=-=Stiller and Konrad, 1999-=-). The motion models are essentially mapping functions that describe the displacement of all points in the image resulting from camera motion. A camera undergoing translation, scale (zoom), and rotati... |

45 |
Segmentation and tracking of piglets in images.
- McFarlane, Schofield
- 1995
(Show Context)
Citation Context ...are non-consecutive, it is unlikely that a non-background outlier will appear in the same location in multiple frames so the median is still a robust estimate. A recursive median estimate is used in (=-=McFarlane and Schofield, 1995-=-); the running estimate of the median is incremented by one if the current observation is greater than the previous estimate, and decremented by one if it is less. Each pixel in B then eventually conv... |

31 |
Background Subtraction Using Markov Thresholds
- Migdal
- 2005
(Show Context)
Citation Context ...n) xii 1Chapter 1 Introduction Background subtraction is a general term for a process which aims to segment moving foreground objects from a relatively stationary background. As it has been noted in (=-=Migdal and Grimson, 2005-=-) and (Radke, 2005), we point out that that there is an important distinction between the background modeling and background detection stages, which comprise the whole subtraction process. As illustra... |

27 | View-based detection and analysis of periodic motion,‖ in
- Cutler, Davis
- 1998
(Show Context)
Citation Context ...n-recursive approach incurs a memory cost to the algorithm since a frame buffer is required. In place of the temporal mean, which is susceptible to outliers, the temporal median is commonly used. In (=-=Cutler and Davis, 1998-=-), the median across 50-200 recent frames is used to estimate the background image, with the assumption that over this span of time, a pixel is occupied by background at least 50% of the time. Alterna... |

21 |
Motion detection and estimation
- Konrad
- 2005
(Show Context)
Citation Context ... to use the so-called phase correlation (PC). This technique is used for block matching. It mitigates exhaustive searching using the correlation criterion as opposed to a pixel-based error criterion (=-=Konrad, 2005-=-). 79 The correlation between two images is denoted C[d] = ∑ n I1[n]I2[n− d] (A.1) To clarify, the correlation is defined over d ∈ Γ ⊆ Λ ⊂ R2 which is a discrete space domain of delays or lags, wherea... |

5 | Real-time adaptive background segmentation
- Butler, Sridharan, et al.
- 2003
(Show Context)
Citation Context ...d memory, since the above process must be done for each pixel and must be reiterated for every frame. To speed up this algorithm, a fast variant dubbed the “group of clusters” method is presented in (=-=Butler et al., 2003-=-). In this variant, the clusters are parameterized by only the weight wk and the centroid ck, which can each be updated quickly. An incoming observation, x, is matched to the cluster whose centroid th... |

4 | Local kernel color histograms for background subtraction
- Noriega, Bascle, et al.
- 2006
(Show Context)
Citation Context ...presented in the literature. Similar in principle to the spatial modeling of (Jodoin et al., 2006a), a technique based on a region descriptor dubbed the “local kernel color histogram” is proposed in (=-=Noriega et al., 2006-=-). In that paper, an image is described by color histograms that are smoothed by a Gaussian kernel, much like the non-parametric approach. Unlike (Jodoin et al., 2006a), however, a spatial kernel is a... |

3 | Light and fast statistical motion detection method based on ergodic model - Jodoin, Mignotte, et al. |

2 | Background subtraction framework based on local spatial distributions - Jodoin, Mignotte, et al. |

1 | Submitted Dec 2006). Distributed detection in sensor networks with limited sensing range - Ermis, Saligrama |

1 | jmmchugh@gmail.com Education Boston University, College of Engineering - Boston, MA M.S. Electrical and Computer Engineering - January 2008 GPA 3.9, Concentration - Signal Processing and Communications Thesis: “Probabilistic Methods for Adaptive Backgroun - Avenue, Allston |