#### DMCA

## 1Adaptive Kalman Filtering for Histogram-based Appearance Learning in Infrared Imagery

### Citations

1005 | An introduction to signal detection and estimation - Poor |

899 | Kernel-based object tracking
- Comaniciu, Ramesh, et al.
(Show Context)
Citation Context ...estimation of the model parameters continuously online [16], or non-parametric, where the target appearance is characterized by empirically derived features that can be updated online during tracking =-=[17]-=-, [18]. Such features may include kernel-based windows [19]–[21], nonparametric or semiparametric contours [22], templates [20], shape descriptors [19], or local statistics [20], [23] including, e.g.,... |

814 | Real-time tracking of non-rigid objects using mean shift
- Comaniciu, Ramesh, et al.
- 2000
(Show Context)
Citation Context ...the foreground/background intensity gA(xk)/gB(xk) and foreground/background local standard deviation (stdev) gC(xk)/gD(xk), which are extracted from yk by using the kernel-based method in [14], [15], =-=[61]-=-. Given xk, the candidate region is characterized by G(xk) defined by G(xk) = {gA(xk),gB(xk),gC(xk),gD(xk)}. (34) A reference target model learned from previous frames is also available that is compos... |

700 | Object tracking: A survey - Yilmaz, Javed, et al. |

356 | Online selection of discriminative tracking features.
- Collins, Liu, et al.
- 2005
(Show Context)
Citation Context ...as discussed previously in [35]–[38]. In these methods, target tracking is performed on an intermediate classification image called a confidence map [35], a likelihood image [36], or a weighted image =-=[37]-=- where each pixel is assigned a probability of belonging to background or foreground. Here we have a different point of view using background for target modeling. Our target model is motivated by the ... |

346 | Robust online appearance models for visual tracking
- Jepson, Fleet, et al.
(Show Context)
Citation Context ...ametric, where a statistical model is typically assumed that captures the key characteristics of the target appearance in a way that facilitates estimation of the model parameters continuously online =-=[16]-=-, or non-parametric, where the target appearance is characterized by empirically derived features that can be updated online during tracking [17], [18]. Such features may include kernel-based windows ... |

331 | Elliptical head tracking using intensity gradients and color histograms
- Birchfield
- 1998
(Show Context)
Citation Context ...tion of the model parameters continuously online [16], or non-parametric, where the target appearance is characterized by empirically derived features that can be updated online during tracking [17], =-=[18]-=-. Such features may include kernel-based windows [19]–[21], nonparametric or semiparametric contours [22], templates [20], shape descriptors [19], or local statistics [20], [23] including, e.g., inten... |

328 | Ensemble tracking”
- Avidan
- 2005
(Show Context)
Citation Context ... has been used to combat the drifting problem involves maintaining explicit appearance models for both the target and the surrounding background. Background information was explicitly incorporated in =-=[35]-=-–[39] to represent the target in terms of features capable of enhancing background discrimination performance. In [14], we proposed a dual foreground-background appearance model comprising four histog... |

278 | Region covariance: A fast descriptor for detection and classification
- Tuzel, Porikli, et al.
- 2006
(Show Context)
Citation Context ...ined structure, the stdev histograms (fC and fD) in all cases were updated using the HS method. We also compare the performance of an alternative target representation using the covariance descriptor =-=[62]-=- that also supports online appearance updates. In addition to the tracking errors, we adopt an overlap metric proposed in [63] to quantify the degree of overlap between the track gate and the actual t... |

207 | Mean-shift blob tracking through scale space,
- Collins
- 2003
(Show Context)
Citation Context ...eter estimation. B. Histogram-based appearance learning Histograms of the pixel intensities have been widely used and were adopted in the appearance models of several recent mean-shift trackers [17], =-=[28]-=-–[30]. Histograms of the local standard deviation (stdev) were also used for mean-shift tracking of IR targets in [23]. The popularity of histogrambased features results at least in part from their si... |

201 | The template update problem. - Matthews, Ishikawa, et al. - 2004 |

127 | Covariance tracking using model update based on lie algebra
- Porikli, Tuzel, et al.
- 2006
(Show Context)
Citation Context ...It was first proposed in [62] for object detection with significant advantages than histogram-based appearance models, and extended to tracking by augmenting with manifold learning-based model update =-=[64]-=-, [65]. In IR tracking, the covariance descriptor involves local intensity, stdev, gradient, orientation and Laplacian information of the target area. Like the other three histogram-based appearance l... |

121 |
Indexing via Color Histograms
- Swain, Ballard
- 1990
(Show Context)
Citation Context ...ns a separate coefficient ξbk for each histogram bin. A. Histogram Similarity Method (HS) In the widely used HS method, the coefficient vector ξk in (1) is updated based on histogram similarity [33], =-=[44]-=-. All Nb entries of ξk share a common value given by the metric ξk = 1− h(fk−1,gk), (2) where h is a normalized histogram similarity measure such as the Bhattacharyya coefficient [17]. In practice, ho... |

118 | Real-time tracking of image regions with changes in geoemtry and illumination. - Hager, Belhumeur - 1996 |

103 |
On the identification of variances and adaptive Kalman filtering
- Mehra
- 1970
(Show Context)
Citation Context ...xploit any observed nonzero correlations at lags other than zero to obtain solutions for the unknown noise variances and/or the optimal gains. Pioneering work in this area was given by Mehra in [41], =-=[49]-=- where the residual autocorrelation was used for adaptive Kalman filtering. Mehra’s method involves a three-step iterative process where a Lyapunov-type equation must be solved at every time step. Und... |

88 | Contour-based object tracking with occlusion handling in video acquired using moblie cameras
- Yilmaz, Li, et al.
- 2004
(Show Context)
Citation Context ...characterized by empirically derived features that can be updated online during tracking [17], [18]. Such features may include kernel-based windows [19]–[21], nonparametric or semiparametric contours =-=[22]-=-, templates [20], shape descriptors [19], or local statistics [20], [23] including, e.g., intensity histograms and their moments. Significant efforts have been directed towards developing methods for ... |

77 | An em-like algorithm for color histogram-based object tracking. In: - Zivkovic, Krose - 2004 |

71 |
Approaches to adaptive filtering
- Mehra
- 1972
(Show Context)
Citation Context ...s to exploit any observed nonzero correlations at lags other than zero to obtain solutions for the unknown noise variances and/or the optimal gains. Pioneering work in this area was given by Mehra in =-=[41]-=-, [49] where the residual autocorrelation was used for adaptive Kalman filtering. Mehra’s method involves a three-step iterative process where a Lyapunov-type equation must be solved at every time ste... |

71 | On the importance of combining wavelet-based nonlinear approximation with coding strategies - Cohen, Daubechies, et al. |

49 |
On-Line Density-Based Appearance Modeling for Object Tracking
- Han, Davis
- 2005
(Show Context)
Citation Context ...utlier components in a Gaussian mixture model (GMM) was proposed in [16], where the model was updated via an expectation maximization (EM) algorithm. GMM-based appearance learning was also applied in =-=[27]-=-, where a mean-shift algorithm was used to update the parameters online. These methods rely on elaborate parametric models and are effective for tracking extended targets with large spatial signatures... |

48 |
Adaptive Kalman filtering for INS/GPS”,
- Mohamed, Shwarz
- 1999
(Show Context)
Citation Context ...us on two different AKF-based appearance learning algorithms that rely on the covariance matching and correlation approaches. C. AKF: Covariance Matching (AKFcov) Covariance matching techniques [41], =-=[47]-=- are based on the relationship that exists between the process and measurement noise variances and the autocorrelation of the innovations process (9). Since the innovations are observable, their autoc... |

37 | Hybrid particle filter and mean shift tracker with adaptive transition model. - Maggio, Cavallaro - 2005 |

34 | Fast Occluded Object Tracking by a Robust Appearance Filter,”
- Nguyen, Smeulders
- 2004
(Show Context)
Citation Context ...l statistics [20], [23] including, e.g., intensity histograms and their moments. Significant efforts have been directed towards developing methods for online appearance learning [1], [3], [16], [24], =-=[25]-=-. For both parametric and non-parametric approaches, the design of an effective learning strategy is strongly coupled to the choice of features. Drift correction strategies for template tracking were ... |

32 | Target Tracking in Airborne Forward Looking Infrared Imagery.
- Yilmaz, Shafique, et al.
- 2000
(Show Context)
Citation Context ... during tracking [17], [18]. Such features may include kernel-based windows [19]–[21], nonparametric or semiparametric contours [22], templates [20], shape descriptors [19], or local statistics [20], =-=[23]-=- including, e.g., intensity histograms and their moments. Significant efforts have been directed towards developing methods for online appearance learning [1], [3], [16], [24], [25]. For both parametr... |

31 | Visual tracking via incremental log-euclidean riemannian subspace learning. In:
- Li, Hu, et al.
- 2008
(Show Context)
Citation Context ... first proposed in [62] for object detection with significant advantages than histogram-based appearance models, and extended to tracking by augmenting with manifold learning-based model update [64], =-=[65]-=-. In IR tracking, the covariance descriptor involves local intensity, stdev, gradient, orientation and Laplacian information of the target area. Like the other three histogram-based appearance learnin... |

30 | About priority encoding transmission - Boucheron, Salamatian - 2000 |

25 | cVehicle tracking using on-line fusion of color and shape features
- Bebis, Miller
- 2004
(Show Context)
Citation Context ...e of an alternative target representation using the covariance descriptor [62] that also supports online appearance updates. In addition to the tracking errors, we adopt an overlap metric proposed in =-=[63]-=- to quantify the degree of overlap between the track gate and the actual target area. Let A and B represent the track gate and the ground-truth bounding box respectively; then the overlap ratio ζ is d... |

21 | Robust track using foreground-background texture discrimination
- Nguyen, Smeulders
- 2006
(Show Context)
Citation Context ...been used to combat the drifting problem involves maintaining explicit appearance models for both the target and the surrounding background. Background information was explicitly incorporated in [35]–=-=[39]-=- to represent the target in terms of features capable of enhancing background discrimination performance. In [14], we proposed a dual foreground-background appearance model comprising four histograms ... |

21 | A new autocovariance leastsquares method for estimating noise covariances. Submitted for publication in Automatica,
- Odelson, Rajamani, et al.
- 2005
(Show Context)
Citation Context ...rst time in an appearance learning application, the unknown process and measurement noise variances are estimated simultaneously using the recently developed autocovariance least-squares (ALS) method =-=[42]-=-, [43]. In order to provide robustness, to accommodate strong ego-motion, and to provide flexibility in dealing with dynamic target size estimation, we adopt a particle filter-based tracker where the ... |

20 |
A recursive multiple model approach to noise identification,”
- Li, Bar-Shalom
- 1994
(Show Context)
Citation Context ...is leads to the adaptive Kalman filter (AKF), which seeks to estimate the unknown noise variances on the fly. A brief overview of AKF methods was given in [41] and more recent surveys appear in [45], =-=[46]-=-. In [41], these techniques were broadly divided into four categories: Bayesian, maximum likelihood (ML), correlation, and covariance matching methods. The Bayesian method requires the evaluation of s... |

14 |
Mean shift blob tracking with kernel histogram filtering and hypothesis testing
- Peng, Yang, et al.
- 2005
(Show Context)
Citation Context ...estimation. B. Histogram-based appearance learning Histograms of the pixel intensities have been widely used and were adopted in the appearance models of several recent mean-shift trackers [17], [28]–=-=[30]-=-. Histograms of the local standard deviation (stdev) were also used for mean-shift tracking of IR targets in [23]. The popularity of histogrambased features results at least in part from their simplic... |

12 | Probabilistic tracking with adaptive feature selection
- Chen, Liu, et al.
(Show Context)
Citation Context ... background for tracking was discussed previously in [35]–[38]. In these methods, target tracking is performed on an intermediate classification image called a confidence map [35], a likelihood image =-=[36]-=-, or a weighted image [37] where each pixel is assigned a probability of belonging to background or foreground. Here we have a different point of view using background for target modeling. Our target ... |

10 |
Adaptive estimation of noise covariance matrices in real-time preprocessing of geophysical data.
- Noriega, Pasupathy
- 1997
(Show Context)
Citation Context ...ce. This leads to the adaptive Kalman filter (AKF), which seeks to estimate the unknown noise variances on the fly. A brief overview of AKF methods was given in [41] and more recent surveys appear in =-=[45]-=-, [46]. In [41], these techniques were broadly divided into four categories: Bayesian, maximum likelihood (ML), correlation, and covariance matching methods. The Bayesian method requires the evaluatio... |

9 |
Identification of optimum filter steady-state gain for systems with unknown noise covariances.
- Carew, Belanger
- 1973
(Show Context)
Citation Context ...p iterative process where a Lyapunov-type equation must be solved at every time step. Under the assumption that the process and measurement noises are wide sense stationary (WSS), Carew and Bélanger =-=[50]-=- developed an improved algorithm that estimates the optimal Kalman gains directly using one matrix inversion and several matrix multiplications, eliminating the need to estimate the process and measur... |

9 | Asymptotically optimal low–complexity sequential lossless coding for piecewise–stationary memoryless sources — part I: The regular case
- Shamir, Costello
- 2000
(Show Context)
Citation Context ...stified by the high frame rate of the imaging sensor compared to the rate at which the target appearance changes. Such assumptions are common, e.g., in the context of audio and video compression [54]–=-=[60]-=-. The nonstationary characteristics of σ2wb(k) and σ2vb(k) directly correlate with the rate at which the target appearance and sensor noise are changing. The piecewise stationary formulation allows us... |

8 | Object tracking with dynamic template update and occlusion detection - Latecki, Miezianko |

8 |
Target tracking in infrared imagery using weighted composite reference function-based decision fusion
- Dawoud, Alam, et al.
- 2006
(Show Context)
Citation Context ... empirically derived features that can be updated online during tracking [17], [18]. Such features may include kernel-based windows [19]–[21], nonparametric or semiparametric contours [22], templates =-=[20]-=-, shape descriptors [19], or local statistics [20], [23] including, e.g., intensity histograms and their moments. Significant efforts have been directed towards developing methods for online appearanc... |

7 |
Automated Tracking and Classification of Infrared Images
- Shaik, Iftekharuddin
(Show Context)
Citation Context ..., or non-parametric, where the target appearance is characterized by empirically derived features that can be updated online during tracking [17], [18]. Such features may include kernel-based windows =-=[19]-=-–[21], nonparametric or semiparametric contours [22], templates [20], shape descriptors [19], or local statistics [20], [23] including, e.g., intensity histograms and their moments. Significant effort... |

7 | Robust visual tracking via pixel classification and integration,”
- Zhang, Rui
- 2006
(Show Context)
Citation Context ...cale and rotation invariance properties [17], [23], [31]. For histogram-based target representations, appearance learning is generally accomplished by iteratively updating a reference histogram [30], =-=[32]-=-, [33]. Typically, the new reference histogram at each iteration is given by a linear weighting of the previous reference histogram and the most recent observation, where the weighting may be based on... |

7 |
Tracking by affine kernel transformations using color and boundary cues
- LEICHTER, LINDENBAUM, et al.
- 2009
(Show Context)
Citation Context ...nd rotation invariance properties [17], [23], [31]. For histogram-based target representations, appearance learning is generally accomplished by iteratively updating a reference histogram [30], [32], =-=[33]-=-. Typically, the new reference histogram at each iteration is given by a linear weighting of the previous reference histogram and the most recent observation, where the weighting may be based on an ap... |

7 | Locally Stationary Covariance and Signal Estimation with Macrotiles
- Donoho, Mallat, et al.
- 2003
(Show Context)
Citation Context ...be justified by the high frame rate of the imaging sensor compared to the rate at which the target appearance changes. Such assumptions are common, e.g., in the context of audio and video compression =-=[54]-=-–[60]. The nonstationary characteristics of σ2wb(k) and σ2vb(k) directly correlate with the rate at which the target appearance and sensor noise are changing. The piecewise stationary formulation allo... |

6 |
Robust object tracking against template drift
- Pan, Hu
- 2007
(Show Context)
Citation Context ...s, making it difficult to maintain both a reliable detection process and a robust track lock over longer time scales – phenomena that have been referred to variously as the “drifting problem” in [1], =-=[2]-=-, the “template update problem” in [3]–[6], and a “stale template condition” in [7]. These challenges are exemplified by the well-known AMCOM closure sequences1 [8]–[15] as well as the newly released ... |

6 | Decision fusion algorithm for target tracking in infrared imagery - Dawoud, Alam, et al. - 2005 |

6 | Target tracking with online feature selection in flir imagery
- Venkataraman, Fan, et al.
- 2007
(Show Context)
Citation Context ...the surrounding background. Background information was explicitly incorporated in [35]–[39] to represent the target in terms of features capable of enhancing background discrimination performance. In =-=[14]-=-, we proposed a dual foreground-background appearance model comprising four histograms that characterize the pixel intensity distribution and the local distribution of the sample stdev over both the t... |

6 | Online updating appearance generative mixture model for Mean-Shift tracking
- Tu, Tao, et al.
(Show Context)
Citation Context ...ng the appearance model when the target has a large spatial extent, they can be susceptible to drifting problems, particularly when applied to smaller targets. Alternatively, a method was proposed in =-=[34]-=- for updating the reference histogram which treats the observed histogram as a realization of a generative model that is a piecewise linear combination of several pairs of histograms computed from rep... |

5 |
Efficient target detection in cluttered FLIR imagery
- Khan, Alam
(Show Context)
Citation Context ...y as the “drifting problem” in [1], [2], the “template update problem” in [3]–[6], and a “stale template condition” in [7]. These challenges are exemplified by the well-known AMCOM closure sequences1 =-=[8]-=-–[15] as well as the newly released SENSIAC ATR dataset.2 One instance of this kind of nonstationary target signature evolution occurs in AMCOM LWIR sequence rng18_17. Here, an LWIR sensor is situated... |

5 | Robust observations for object tracking
- Han, Davis
- 2005
(Show Context)
Citation Context ...We present a target model that involves the local statistics of both the target and its surrounding background, as shown in Fig. 5. The use of background for tracking was discussed previously in [35]–=-=[38]-=-. In these methods, target tracking is performed on an intermediate classification image called a confidence map [35], a likelihood image [36], or a weighted image [37] where each pixel is assigned a ... |

5 |
Comments on “Identification of optimum filter steadystate gain for systems with unknown noise covariances”.
- Neethling, Young
- 1974
(Show Context)
Citation Context ...the need to estimate the process and measurement noise variances explicitly and avoiding the requirement to iteratively solve the Lyapunov equation associated with Mehra’s method. Neethling and Young =-=[51]-=- introduced a related weighted least squares technique that improves the statistical efficiency of the methods in [41], [49], [50] and incorporates a side constraint to guarantee positive semi-definit... |

4 |
A drifting-proof framework for tracking and online appearance learning
- Han, Liu, et al.
- 2007
(Show Context)
Citation Context ...scales, making it difficult to maintain both a reliable detection process and a robust track lock over longer time scales – phenomena that have been referred to variously as the “drifting problem” in =-=[1]-=-, [2], the “template update problem” in [3]–[6], and a “stale template condition” in [7]. These challenges are exemplified by the well-known AMCOM closure sequences1 [8]–[15] as well as the newly rele... |

4 | A novel multiple tracking system for UAV platforms - Yi, Zhang |

4 |
Adaptive Tracking of Multiple Hot-Spot Target IR Images
- Maybeck, Rodgers
(Show Context)
Citation Context ...r local statistics [20], [23] including, e.g., intensity histograms and their moments. Significant efforts have been directed towards developing methods for online appearance learning [1], [3], [16], =-=[24]-=-, [25]. For both parametric and non-parametric approaches, the design of an effective learning strategy is strongly coupled to the choice of features. Drift correction strategies for template tracking... |

4 | Estimating Disturbance Covariances From Data for Improved Control Performance.
- Odelson
- 2003
(Show Context)
Citation Context ...riance estimates are more stable than those delivered by Mehra’s method and converge asymptotically to the optimal values with increasing sample size. However, the proof of convergence given in [42], =-=[52]-=- depends explicitly on assumptions that the system is time invariant and that the process and measurement noises are WSS (extension to a time varying system with WSS noises was given in [53]). The ALS... |

3 |
Extended target tracking using projection curves and matching pel count,”Opt
- Peng, Zhang, et al.
- 2007
(Show Context)
Citation Context ...reliable detection process and a robust track lock over longer time scales – phenomena that have been referred to variously as the “drifting problem” in [1], [2], the “template update problem” in [3]–=-=[6]-=-, and a “stale template condition” in [7]. These challenges are exemplified by the well-known AMCOM closure sequences1 [8]–[15] as well as the newly released SENSIAC ATR dataset.2 One instance of this... |

3 | Low bit rate vector quantization of outlier contaminated data based on shells of Golay codecs - Tabus, Vasiache |

3 | Joint fixed-rate universal lossy coding and identification of continuous-alphabet memoryless sources - Raginsky - 2008 |

2 | Online consistency checking for AM-FM target tracks - Mould, Nguyen, et al. |

2 | Appearance learning by adaptive Kalman filters for FLIR tracking
- Venkataraman, Fan, et al.
- 2009
(Show Context)
Citation Context ... the “drifting problem” in [1], [2], the “template update problem” in [3]–[6], and a “stale template condition” in [7]. These challenges are exemplified by the well-known AMCOM closure sequences1 [8]–=-=[15]-=- as well as the newly released SENSIAC ATR dataset.2 One instance of this kind of nonstationary target signature evolution occurs in AMCOM LWIR sequence rng18_17. Here, an LWIR sensor is situated on a... |

2 | Target tracking in infrared image sequences using diverse AdaBoostSVM
- Wang, Wu, et al.
(Show Context)
Citation Context ...non-parametric, where the target appearance is characterized by empirically derived features that can be updated online during tracking [17], [18]. Such features may include kernel-based windows [19]–=-=[21]-=-, nonparametric or semiparametric contours [22], templates [20], shape descriptors [19], or local statistics [20], [23] including, e.g., intensity histograms and their moments. Significant efforts hav... |

2 | Tracking through changes in scale
- Lankton, Nakhmani, et al.
- 2008
(Show Context)
Citation Context ...tric and non-parametric approaches, the design of an effective learning strategy is strongly coupled to the choice of features. Drift correction strategies for template tracking were proposed in [3], =-=[26]-=-. A more sophisticated model combining stable, wandering, and outlier components in a Gaussian mixture model (GMM) was proposed in [16], where the model was updated via an expectation maximization (EM... |

2 |
Application of a new data-based covariance estimation technique to a nonlinear industrial blending drum,” Texas-Winsconsin Modeling and Control Consortium
- Rajamani, Rawlings, et al.
- 2007
(Show Context)
Citation Context ... in [42], [52] depends explicitly on assumptions that the system is time invariant and that the process and measurement noises are WSS (extension to a time varying system with WSS noises was given in =-=[53]-=-). The ALS algorithm in [42] is primarily meant for identifying the system noise properties in an offline learning process under WSS assumptions. First, the filter innovations are obtained from the ob... |

2 | Speech data compression through sparse coding of innovations - Ramabadran, Sinha - 1994 |

1 | Dual domain auxiliary particle filter with integrated target signature update
- Johnston, Mould, et al.
- 2009
(Show Context)
Citation Context ...rack lock over longer time scales – phenomena that have been referred to variously as the “drifting problem” in [1], [2], the “template update problem” in [3]–[6], and a “stale template condition” in =-=[7]-=-. These challenges are exemplified by the well-known AMCOM closure sequences1 [8]–[15] as well as the newly released SENSIAC ATR dataset.2 One instance of this kind of nonstationary target signature e... |

1 | Robust automatic target tracking based on a Bayesian ego-motion compensation framework for airborne FLIR imagery - Blanco, Jaureguizar, et al. |

1 |
der Boomagaard, “Occlusion robust adaptive template tracking
- Nguyen, Worring, et al.
(Show Context)
Citation Context ...ackground structure and fail. Improved histogram estimation was achieved by modeling the temporal evolution of the reference histogram in an adaptive Kalman filtering (AKF) framework in [30]. In [2], =-=[40]-=-, the AKF measurement noise variance was estimated from the first frame and was assumed stationary, while the process noise variance was estimated online using covariance matching [41]. A robust Kalma... |