Back

Color-Based Probabilistic Tracking

P. PérezC. HueJ. VermaakMichel Gangnet

article2002ECCV1,577 citations

Proposes a particle filter tracking framework based on color histogram matching that tracks multiple posterior modes to reliably recover targets amidst background clutter, temporary occlusions, and severe shape deformations.

Listen

Visual tracking systems often fail when monitored objects change shape, experience motion blur, or face temporary visual blockages. Standard deterministic color trackers provide low-cost tracking by matching target color histograms across frames, but they easily lose targets in cluttered backgrounds or during complete occlusions. The article addresses these limitations by developing and evaluating a probabilistic visual tracking framework that combines color histogram matching with sequential Monte Carlo estimation, commonly known as particle filtering.

The evaluated approach uses a particle filter to propagate multiple plausible target locations simultaneously across frames. The tracker models candidate image patches in the Hue-Saturation-Value color space, evaluates their similarity against a reference color model using the Bhattacharyya distance, and estimates the target position and scale over time. The framework was evaluated across video sequences with challenging conditions such as fast movement, background color distractions, occlusions, and multiple intersecting objects.

The key findings show that the probabilistic approach overcomes the critical failure modes of deterministic color trackers. First, propagating multiple hypotheses prevents the tracker from locking onto distracting background colors and allows it to fully recover target trajectories after complete occlusions lasting multiple frames. Second, splitting the target window into a multi-part color model preserves coarse spatial layouts, which eliminates tracking drift and improves scale estimation. Third, incorporating static background models and skin-detection initialization enables automated, real-time multi-object tracking, maintaining individual identities without identity swapping when targets cross paths. Computationally, a non-optimized implementation achieved 50 frames per second using 100 particles on a 747 MHz processor for 25 by 25 pixel targets.

These results demonstrate that probabilistic color tracking delivers high robustness at low computational cost, making it suitable for real-time video surveillance, user interfaces, and automated video editing without requiring specialized hardware. The framework is immediately actionable for applications in fixed-camera environments and multi-target tracking. Organizations implementing this system should adopt multi-part color models to reduce drift and use background models where stationary cameras are available. Further research should explore automated region splitting, adaptive parameter tuning, and combining color likelihoods with edge or contour information to handle complex, evolving scenes.

IRISAMicrosoftcolor histogramcomputational efficiencydeterministiccolor-based probabilistic trackingparticle filteringreference color histogram
Pérez et al (2002).pdf
  • Paper: CONDENSATION—Conditional Density Propagation for Visual Tracking, MICHAEL ISARD et al. (1998). It introduces the CONDENSATION particle filtering algorithm for visual tracking that serves as the probabilistic framework for propagating target state distributions in the source.
  • Paper: Color indexing, Michael J. Swain et al. (1991). It establishes color histogram matching and intersection metrics for real-time object identification and localization upon which the source's color-based tracking relies.
  • Paper: Learning Patterns of Activity Using Real-Time Tracking, Chris Stauffer et al. (2000). It develops adaptive background subtraction using mixture-of-Gaussians modeling, providing the foundational background modeling integrated by the source for real-time tracking.
  • Paper: Probabilistic Visual Learning for Object Representation, B. Moghaddam et al. (1997). It lays the groundwork for probabilistic density estimation and adaptive statistical modeling for visual tracking in dynamic environments.
  • Paper: Non-parametric Model for Background Subtraction, Ahmed Elgammal et al. (2000). It provides non-parametric background subtraction and chromaticity analysis that inform probabilistic foreground-background modeling.
  • Paper: Kernel-Based Object Tracking, Dorin Comaniciu et al. (2003). It introduces kernel-based mean-shift tracking with Bhattacharyya distance metrics, offering an alternative deterministic gradient-ascent approach to histogram-based tracking.
  • Paper: Incremental Learning for Robust Visual Tracking, David A. Ross et al. (2008). It advances particle-filter-based visual tracking by incorporating incremental subspace learning to handle appearance and illumination variations dynamically.
  • Paper: Staple: Complementary Learners for Real-Time Tracking, Luca Bertinetto et al. (2015). It extends color-based tracking concepts by fusing dense color histograms with spatial correlation filter templates in a complementary real-time framework.
  • Paper: Ensemble Tracking, S. Avidan (2005). It builds beyond simple appearance modeling by framing tracking as an adaptive online classification problem distinguishing target from background.
  • Paper: Robust Object Tracking with Online Multiple Instance Learning, Boris Babenko et al. (2011). It advances discriminative visual tracking under ambiguous target bounding boxes by introducing online multiple instance learning to mitigate drift.
  • Paper: Tracking-Learning-Detection, Zdenek Kalal et al. (2012). It develops a complete long-term tracking framework combining frame-to-frame tracking, redetection, and learning to handle persistent target disappearance.
  • Paper: Object Tracking Benchmark, Yi Wu et al. (2015). It provides a comprehensive standardized benchmark and evaluation protocol to systematically evaluate online visual tracking paradigms.
Cover for Color-Based Probabilistic Tracking

Abstract

Color-based trackers recently proposed in [3,4,5] have been proved robust and versatile for a modest computational cost. They are especially appealing for tracking tasks where the spatial structure of the tracked objects exhibits such a dramatic variability that trackers based on a space-dependent appearance reference would break down very fast. Trackers in [3,4,5] rely on the deterministic search of a window whose color content matches a reference histogram color model.

Relying on the same principle of color histogram distance, but within a probabilistic framework, we introduce a new Monte Carlo tracking technique. The use of a particle filter allows us to better handle color clutter in the background, as well as complete occlusion of the tracked entities over a few frames.

This probabilistic approach is very flexible and can be extended in a number of useful ways. In particular, we introduce the following ingredients: multi-part color modeling to capture a rough spatial layout ignored by global histograms, incorporation of a background color model when relevant, and extension to multiple objects.

Topics

color histogramcomputational efficiencydeterministicframestracked objectsobject appearancetracking algorithmclutter background

Keywords

color-based probabilistic trackingparticle filteringreference color histogrambackground subtractionMonte Carlo trackingface trackingcolor distributionKalman filterGaussian mixture

Table of Contents

  • 1 Introduction
  • 2 Probabilistic Tracking
  • 2.1 Sequential Monte Carlo Tracking
  • 2.2 State Space and Dynamics
  • 2.3 Color Model
  • 3 Results and Extensions
  • 3.1 Base Tracker
  • 3.2 Multi-part Color Model
  • 3.3 Background Modeling
  • 3.4 Multiple Objects
  • 3.5 Automatic Initialization on Skin
  • 4 Discussion
  • References
  • Appendix: A Connection with Bramble [11]

Knowls

  1. Knowl 1 — Color-Based Particle Filter Tracking Algorithm

    algorithm

    The single-object color-based particle filter recursively estimates the posterior distribution p(xt∣y0:t)p(\mathbf{x}_t \mid \mathbf{y}_{0:t}) of target state xt\mathbf{x}_t given video frames y0:t\mathbf{y}_{0:t} using a set of MM weighted samples (particles).

    Input: Reference color histogram q∗={q∗(n)}n=1N\mathbf{q}^* = \{q^*(n)\}_{n=1}^N, particle set at time tt {xtm}m=1M\{\mathbf{x}_t^m\}_{m=1}^M, parameter λ\lambda
    Output: Estimated state x^t+1\hat{\mathbf{x}}_{t+1}, resampled particle set {xt+1m}m=1M\{\mathbf{x}_{t+1}^m\}_{m=1}^M
    for m=1m = 1 to MM do
        Draw predicted state x~t+1m\tilde{\mathbf{x}}_{t+1}^m from autoregressive dynamics p(xt+1∣xtm)p(\mathbf{x}_{t+1} \mid \mathbf{x}_t^m)
        Extract candidate color histogram qt+1(x~t+1m)\mathbf{q}_{t+1}(\tilde{\mathbf{x}}_{t+1}^m) from frame yt+1\mathbf{y}_{t+1} within region R(x~t+1m)R(\tilde{\mathbf{x}}_{t+1}^m)
        Compute distance D[q∗,qt+1(x~t+1m)]=1−∑n=1Nq∗(n)qt+1(n;x~t+1m)D[\mathbf{q}^*, \mathbf{q}_{t+1}(\tilde{\mathbf{x}}_{t+1}^m)] = \sqrt{1 - \sum_{n=1}^N \sqrt{q^*(n) q_{t+1}(n; \tilde{\mathbf{x}}_{t+1}^m)}}
        Compute unnormalized weight wt+1m=exp⁡(−λD2[q∗,qt+1(x~t+1m)])w_{t+1}^m = \exp(-\lambda D^2[\mathbf{q}^*, \mathbf{q}_{t+1}(\tilde{\mathbf{x}}_{t+1}^m)])
    end for
    Compute normalization factor W=∑k=1Mwt+1kW = \sum_{k=1}^M w_{t+1}^k
    for m=1m = 1 to MM do
        Set normalized weight πt+1m=wt+1m/W\pi_{t+1}^m = w_{t+1}^m / W
    end for
    for m=1m = 1 to MM do
        Sample index a(m)∈{1,…,M}a(m) \in \{1, \dots, M\} with probability πt+1a(m)\pi_{t+1}^{a(m)}
        Set resampled particle xt+1m=x~t+1a(m)\mathbf{x}_{t+1}^m = \tilde{\mathbf{x}}_{t+1}^{a(m)}
    end for
    Compute state estimate x^t+1=1M∑m=1Mxt+1m\hat{\mathbf{x}}_{t+1} = \frac{1}{M} \sum_{m=1}^M \mathbf{x}_{t+1}^m
    return x^t+1,{xt+1m}m=1M\hat{\mathbf{x}}_{t+1}, \{\mathbf{x}_{t+1}^m\}_{m=1}^M

    In typical tracking runs, M=100M = 100 particles are used with λ=20\lambda = 20.

  2. Knowl 2 — Bhattacharyya-Based Color Observation Likelihood Model

    model/method

    Given a reference color histogram q∗={q∗(n)}n=1N\mathbf{q}^* = \{q^*(n)\}_{n=1}^N with ∑n=1Nq∗(n)=1\sum_{n=1}^N q^*(n) = 1 and a candidate color histogram qt(xt)={qt(n;xt)}n=1N\mathbf{q}_t(\mathbf{x}_t) = \{q_t(n; \mathbf{x}_t)\}_{n=1}^N extracted over the image region R(xt)R(\mathbf{x}_t) defined by state xt\mathbf{x}_t, the distance D[q∗,qt(xt)]D[\mathbf{q}^*, \mathbf{q}_t(\mathbf{x}_t)] derived from the Bhattacharyya coefficient is defined as:

    D[q∗,qt(xt)]=[1−∑n=1Nq∗(n)qt(n;xt)]12D[\mathbf{q}^*, \mathbf{q}_t(\mathbf{x}_t)] = \left[ 1 - \sum_{n=1}^N \sqrt{q^*(n) q_t(n; \mathbf{x}_t)} \right]^{\frac{1}{2}}

    This metric satisfies the properties of a proper distance, is bounded in [0,1][0, 1], and does not diverge on empty histogram bins.

    Assuming the squared Bhattacharyya distance D2D^2 exhibits an exponential distribution, the observation likelihood p(yt∣xt)p(\mathbf{y}_t \mid \mathbf{x}_t) of the image frame yt\mathbf{y}_t given state xt\mathbf{x}_t is defined as:

    p(yt∣xt)∝exp⁡(−λD2[q∗,qt(xt)])p(\mathbf{y}_t \mid \mathbf{x}_t) \propto \exp\left(-\lambda D^2[\mathbf{q}^*, \mathbf{q}_t(\mathbf{x}_t)]\right)

    where λ\lambda is a positive weighting parameter (fixed empirically to λ=20\lambda = 20).

  3. Knowl 3 — Multi-Part Spatial Color Likelihood Model

    model/method

    To preserve coarse spatial color layout without introducing additional transformation parameters, the target bounding region R(x)R(\mathbf{x}) defined by state x\mathbf{x} is partitioned into JJ non-overlapping sub-regions:

    R(x)=⋃j=1JRj(x)R(\mathbf{x}) = \bigcup_{j=1}^J R_j(\mathbf{x})

    Each sub-region j∈{1,…,J}j \in \{1, \dots, J\} is associated with its own reference color histogram qj∗={qj∗(n)}n=1N\mathbf{q}_j^* = \{q_j^*(n)\}_{n=1}^N. Assuming conditional independence of image observations across the sub-regions given state xt\mathbf{x}_t, the composite data likelihood is defined as:

    p(yt∣xt)∝exp⁡(−λ∑j=1JD2[qj∗,qj,t(xt)])p(\mathbf{y}_t \mid \mathbf{x}_t) \propto \exp\left(-\lambda \sum_{j=1}^J D^2[\mathbf{q}_j^*, \mathbf{q}_{j,t}(\mathbf{x}_t)]\right)

    where qj,t(xt)\mathbf{q}_{j,t}(\mathbf{x}_t) is the color histogram computed within sub-region Rj(xt)R_j(\mathbf{x}_t) in frame yt\mathbf{y}_t, λ\lambda is the scaling factor, and D[⋅,⋅]D[\cdot, \cdot] is the Bhattacharyya distance:

    D[qj∗,qj,t(xt)]=[1−∑n=1Nqj∗(n)qj,t(n;xt)]12D[\mathbf{q}_j^*, \mathbf{q}_{j,t}(\mathbf{x}_t)] = \left[ 1 - \sum_{n=1}^N \sqrt{q_j^*(n) q_{j,t}(n; \mathbf{x}_t)} \right]^{\frac{1}{2}}

  4. Knowl 4 — Multi-Object Observation Likelihood with Depth Marginalization

    model/method

    When tracking kk objects simultaneously, the joint state vector is the concatenation of individual object states xt=(x1,t,…,xk,t)\mathbf{x}_t = (\mathbf{x}_{1,t}, \dots, \mathbf{x}_{k,t}), where object ii has target reference histogram qi∗\mathbf{q}_i^*. If bounding regions R(xi,t)R(\mathbf{x}_{i,t}) do not overlap, the joint likelihood is the product of individual object likelihoods.

    When two objects ii and jj overlap (R(xi,t)∩R(xj,t)≠∅R(\mathbf{x}_{i,t}) \cap R(\mathbf{x}_{j,t}) \neq \emptyset), to prevent double-counting pixels, the likelihood marginalizes over the two equally likely relative depth hypotheses (object ii occluding jj, and object jj occluding ii):

    p(yt∣xi,t,xj,t)=0.5[pij(yt∣xi,t,xj,t)+pji(yt∣xi,t,xj,t)]p(\mathbf{y}_t \mid \mathbf{x}_{i,t}, \mathbf{x}_{j,t}) = 0.5 \left[ p_{ij}(\mathbf{y}_t \mid \mathbf{x}_{i,t}, \mathbf{x}_{j,t}) + p_{ji}(\mathbf{y}_t \mid \mathbf{x}_{i,t}, \mathbf{x}_{j,t}) \right]

    where pij(yt∣xi,t,xj,t)p_{ij}(\mathbf{y}_t \mid \mathbf{x}_{i,t}, \mathbf{x}_{j,t}) is the likelihood under the hypothesis that object ii occludes object jj:

    pij(yt∣xi,t,xj,t)∝exp⁡(−λ[D2[qi∗,qt(xi,t)]+D2[qj∗,qt(xj,t∣xi,t)]])p_{ij}(\mathbf{y}_t \mid \mathbf{x}_{i,t}, \mathbf{x}_{j,t}) \propto \exp\left(-\lambda \left[ D^2[\mathbf{q}_i^*, \mathbf{q}_t(\mathbf{x}_{i,t})] + D^2[\mathbf{q}_j^*, \mathbf{q}_t(\mathbf{x}_{j,t} \mid \mathbf{x}_{i,t})] \right]\right)

    Here, qt(xj,t∣xi,t)\mathbf{q}_t(\mathbf{x}_{j,t} \mid \mathbf{x}_{i,t}) denotes the color histogram gathered exclusively over the visible non-occluded subregion R(xj,t)∖R(xi,t)R(\mathbf{x}_{j,t}) \setminus R(\mathbf{x}_{i,t}).

  5. Knowl 5 — Background-Aware Color Observation Likelihood

    model/method

    When tracking in scenes observed by a static camera where an offline reference background image y~\tilde{\mathbf{y}} is available, the likelihood can penalize similarity to the co-located background region. Let q∗\mathbf{q}^* denote the reference foreground histogram, qt(x)\mathbf{q}_t(\mathbf{x}) the candidate histogram gathered in region R(x)R(\mathbf{x}) of frame yt\mathbf{y}_t, and q∙(x)\mathbf{q}^\bullet(\mathbf{x}) the background histogram extracted over the exact same spatial region R(x)R(\mathbf{x}) in the reference background image y~\tilde{\mathbf{y}}.

    The background-aware data likelihood is defined as:

    p(yt∣xt)∝exp⁡(−λ[D2[q∗,qt(xt)]−D2[q∙(xt),qt(xt)]])p(\mathbf{y}_t \mid \mathbf{x}_t) \propto \exp\left(-\lambda \left[ D^2[\mathbf{q}^*, \mathbf{q}_t(\mathbf{x}_t)] - D^2[\mathbf{q}^\bullet(\mathbf{x}_t), \mathbf{q}_t(\mathbf{x}_t)] \right]\right)

    where D[⋅,⋅]D[\cdot, \cdot] is the Bhattacharyya distance and λ\lambda is a positive parameter (lambda=20\\lambda = 20). The term −D2[q∙(xt),qt(xt)]-D^2[\mathbf{q}^\bullet(\mathbf{x}_t), \mathbf{q}_t(\mathbf{x}_t)] rewards hypotheses whose color distribution diverges from the static background.

  6. Knowl 6 — Decoupled Chromatic and Achromatic HSV Color Histogram Representation

    model/method

    Color histograms are computed in Hue-Saturation-Value (HSV) space to decouple chromaticity from illumination and shading variations. The complete histogram consists of N=NhNs+NvN = N_h N_s + N_v total bins:

    1. Chromatic pixels whose saturation S>0.1S > 0.1 and value V>0.2V > 0.2 are assigned to a two-dimensional Hue-Saturation grid of Nh×NsN_h \times N_s bins.
    2. Achromatic / low-reliability pixels (S≤0.1S \le 0.1 or V≤0.2V \le 0.2) are mapped to NvN_v value-only (intensity) bins.

    For a hypothesized region R(x)R(\mathbf{x}), the normalized histogram qt(x)={qt(n;x)}n=1N\mathbf{q}_t(\mathbf{x}) = \{q_t(n; \mathbf{x})\}_{n=1}^N is computed using uniform bin counting (w≡1w \equiv 1):

    qt(n;x)=1∣R(x)∣∑u∈R(x)δ[bt(u)−n]q_t(n; \mathbf{x}) = \frac{1}{|R(\mathbf{x})|} \sum_{\mathbf{u} \in R(\mathbf{x})} \delta[b_t(\mathbf{u}) - n]

    where u\mathbf{u} indexes pixel grid locations within region R(x)R(\mathbf{x}), bt(u)∈{1,…,N}b_t(\mathbf{u}) \in \{1, \dots, N\} denotes the bin index of pixel color yt(u)\mathbf{y}_t(\mathbf{u}), and δ\delta is the Kronecker delta function. Standard settings are Nh=Ns=Nv=10N_h = N_s = N_v = 10, giving N=110N = 110 bins.

  7. Knowl 7 — Second-Order Autoregressive State Dynamics for Bounding Region Tracking

    model/method

    The target region in frame tt is parametrized by a transformation applied to a base zero-centered window WW: R(xt)=dt+stWR(\mathbf{x}_t) = \mathbf{d}_t + s_t W, where dt=(xt,yt)\mathbf{d}_t = (x_t, y_t) is the 2D pixel translation and sts_t is the scale factor.

    The hidden state is defined using current and preceding state parameters as xt=(dt,dt−1,st,st−1)\mathbf{x}_t = (\mathbf{d}_t, \mathbf{d}_{t-1}, s_t, s_{t-1}). The temporal evolution follows a second-order autoregressive (AR) dynamic model:

    xt+1=Axt+Bxt−1+Cvt,vt∼N(0,Σ)\mathbf{x}_{t+1} = A \mathbf{x}_t + B \mathbf{x}_{t-1} + C \mathbf{v}_t, \quad \mathbf{v}_t \sim \mathcal{N}(0, \Sigma)

    where the dynamics decomposes into three independent constant-velocity processes on xtx_t, yty_t, and sts_t. The noise covariance Σ\Sigma is diagonal with standard deviations σx=1 pixel/frame\sigma_x = 1\text{ pixel/frame}, σy=1 pixel/frame\sigma_y = 1\text{ pixel/frame}, and σs=0.1 frame−1\sigma_s = 0.1\text{ frame}^{-1}.

  8. Knowl 8 — Appearance and Motion-Based Automatic Track Initialization

    model/method

    Automatic track initialization for human faces combines an offline-learned skin color distribution with temporal motion detection:

    1. Skin Detection: A normalized HSV histogram q~\tilde{\mathbf{q}} with N=NhNs+NvN = N_h N_s + N_v bins is learned offline from labeled face images. Pixels u\mathbf{u} in frame yt\mathbf{y}_t are classified as skin if their likelihood q~[bt(u)]\tilde{q}[b_t(\mathbf{u})] exceeds a threshold (e.g., 0.30.3).
    2. Motion Sieve: In stationary camera setups, background false alarms are filtered out by retaining only pixels whose absolute frame-difference ∣yt(u)−yt−1(u)∣|\mathbf{y}_t(\mathbf{u}) - \mathbf{y}_{t-1}(\mathbf{u})| exceeds a motion threshold (e.g., 1010).
    3. Track Instantiation: Connected components of motion-filtered skin pixels that match the expected aspect ratio and size range of the target, and do not lie near existing particle clouds, instantiate a new tracked object xi,t\mathbf{x}_{i,t}.

  9. Knowl 9 — Point-Wise Likelihood Ratio Equivalence of Histogram Divergence

    theoretical result

    When a histogram-based observation likelihood is defined via the Kullback-Leibler (KL) divergence instead of the squared Bhattacharyya distance, it is mathematically equivalent to the product of point-wise data likelihood ratios over the pixels u∈R(x)\mathbf{u} \in R(\mathbf{x}):

    exp⁡(−λKL[qt(x)∥q∗])≈∏u∈R(x)(q∗[bt(u)]qt[bt(u);x])λ∣R(x)∣\exp\left(-\lambda \text{KL}[\mathbf{q}_t(\mathbf{x}) \parallel \mathbf{q}^*]\right) \approx \prod_{\mathbf{u} \in R(\mathbf{x})} \left( \frac{q^*[b_t(\mathbf{u})]}{q_t[b_t(\mathbf{u}); \mathbf{x}]} \right)^{\frac{\lambda}{|R(\mathbf{x})|}}

    where KL[qt(x)∥q∗]=∑n=1Nqt(n;x)log⁡qt(n;x)q∗(n)\text{KL}[\mathbf{q}_t(\mathbf{x}) \parallel \mathbf{q}^*] = \sum_{n=1}^N q_t(n; \mathbf{x}) \log \frac{q_t(n; \mathbf{x})}{q^*(n)}, bt(u)b_t(\mathbf{u}) is the color bin at pixel u\mathbf{u}, and ∣R(x)∣|R(\mathbf{x})| is the pixel count in region R(x)R(\mathbf{x}).

    When a static background model q∙(x)\mathbf{q}^\bullet(\mathbf{x}) is incorporated, the divergence difference similarly reduces to:

    exp⁡(−λ(KL[qt(x)∥q∗]−KL[qt(x)∥q∙(x)]))≈∏u∈R(x)(q∗[bt(u)]q∙[bt(u);x])λ∣R(x)∣\exp\left(-\lambda \left( \text{KL}[\mathbf{q}_t(\mathbf{x}) \parallel \mathbf{q}^*] - \text{KL}[\mathbf{q}_t(\mathbf{x}) \parallel \mathbf{q}^\bullet(\mathbf{x})] \right)\right) \approx \prod_{\mathbf{u} \in R(\mathbf{x})} \left( \frac{q^*[b_t(\mathbf{u})]}{q^\bullet[b_t(\mathbf{u}); \mathbf{x}]} \right)^{\frac{\lambda}{|R(\mathbf{x})|}}

    which matches the form of pixel-level likelihood ratios between foreground and background observation models.

Coverage note — Deliberately omitted qualitative comparison image figures and standard tutorial summaries of sequential Monte Carlo filtering.

References

  1. 1.S.T. Birchfield. Elliptical head tracking using intensity gradients and color histograms. In Proc. Conf. Comp. Vision Pattern Rec., pages 232–237, Santa Barbara, CA, June 1998.
  2. 2.M. Black and A. Jepson. Eigentracking: Robust matching and tracking of articulated objects using a view-based representation. Int. J. Computer Vision, 26(1):63–84, 1998.
  3. 3.G.R. Bradski. Computer vision face tracking as a component of a perceptual user interface. In Workshop on Applications of Computer Vision, pages 214–219, Princeton, NJ, Oct. 1998.
  4. 4.H.T. Chen and T.L. Liu. Trust-region methods for real-time tracking. In Proc. Int. Conf. Computer Vision, pages II: 717–722, Vancouver, Canada, July 2001.
  5. 5.D. Comaniciu, V. Ramesh, and P. Meer. Real-time tracking of non-rigid objects using mean shift. In Proc. Conf. Comp. Vision Pattern Rec., pages II:142–149, Hilton Head, SC, June 2000.
  6. 6.A. Doucet, S. Godsill, and C. Andrieu. On sequential Monte Carlo sampling methods for Bayesian filtering. Statistics and Computing, 10(3):197–208, 2000.
  7. 7.J. Foley, A Van Dam, S. Feiner, and J. Hughes. Computer Graphics: Principles and Practice. Addison-Wesley, 1990.
  8. 8.D. Gavrila. Pedestrian detection from a moving vehicle. In Proc. Europ. Conf. Computer Vision, Dublin, Ireland, June 2000.
  9. 9.G.D. Hager and P.N. Belhumeur. Efficient region tracking with parametric models of geometry and illumination. IEEE Trans. Pattern Anal. Machine Intell., 20(10):1025–1039, 1998.
  10. 10.M. Isard and A. Blake. Condensation–conditional density propagation for visual tracking. Int. J. Computer Vision, 29(1):5–28, 1998.
  11. 11.M. Isard and J. MacCormick. BraMBLe: a Bayesian multiple-blob tracker. In Proc. Int. Conf. Computer Vision, pages II: 34–41, Vancouver, Canada, July 2001.
  12. 12.C. Kervrann and F. Heitz. A hierarchical Markov modeling approach for the segmentation and tracking of deformable shape. Graph. Mod. Image Proc., 60(3):173–195, 1998.
  13. 13.J. MacCormick and A. Blake. A probabilistic exclusion principle for tracking multiple objects. In Proc. Int. Conf. Computer Vision, pages 572–578, 1999.
  14. 14.H.T. Nguyen, M. Worring, and R. van den Boomgaard. Occlusion robust adaptive template tracking. In Proc. Int. Conf. Computer Vision, pages I: 678–683, Vancouver, Canada, July 2001.
  15. 15.N.P. Papanikolopoulos, P.K. Khosla, and T. Kanade. Visual tracking of a moving target by a camera mounted on a robot: A combination of control and vision. IEEE Trans. Robotics and Automation, 9:14–35, 1993.
  16. 16.N. Peterfreund. The velocity snake: Deformable contour for tracking in spatio-velocity space. Computer Vision and Image Understanding, 73(3):346–356, 1999.
  17. 17.A. Rahimi, L.P. Morency, and T. Darrell. Reducing drift in parametric motion tracking. In Proc. Int. Conf. Computer Vision, pages I: 315–322, Vancouver, Canada, July 2001.
  18. 18.M. Spengler and B. Schiele. Towards robust multi-cue integration for visual tracking. In Int. Workshop on Computer Vision Systems, Vancouver, Canada, July 2001.
  19. 19.J. Vermaak, P. P´erez, M. Gangnet, and A. Blake. Towards improved observation models for visual tracking: selective adaptation. In Proc. Europ. Conf. Computer Vision, Copenhagen, Denmark, May 2002.
  20. 20.Y. Wu and T. Huang. Color tracking by transductive learning. In Proc. Conf. Comp. Vision Pattern Rec., pages I:133–138, Hilton Head, SC, June 2000.
  21. 21.Y. Wu and T.S. Huang. A co-inference approach to robust visual tracking. In Proc. Int. Conf. Computer Vision, pages II: 26–33, Vancouver, Canada, July 2001.

Citation

MLA
Pérez, P., et al. “Color-Based Probabilistic Tracking”. Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2002, pp. 661–75, https://doi.org/10.1007/3-540-47969-4_44.
APA
Pérez, P., Hue, C., Vermaak, J., & Gangnet, M. (2002). Color-Based Probabilistic Tracking. In Lecture Notes in Computer Science (pp. 661–675). Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-47969-4_44
Chicago
Pérez, P., C. Hue, J. Vermaak, and M. Gangnet. 2002. “Color-Based Probabilistic Tracking”. In Lecture Notes in Computer Science. Springer Berlin Heidelberg. https://doi.org/10.1007/3-540-47969-4_44.
Harvard
Pérez, P. et al. (2002) “Color-Based Probabilistic Tracking”, Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 661–675. Available at: https://doi.org/10.1007/3-540-47969-4_44.
Vancouver
1. Pérez P, Hue C, Vermaak J, Gangnet M (2002) Color-Based Probabilistic Tracking. In: Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp 661–675

BibTeX

@inbook{P_rez_2002, title={Color-Based Probabilistic Tracking}, ISBN={9783540479697}, ISSN={0302-9743}, url={http://dx.doi.org/10.1007/3-540-47969-4_44}, DOI={10.1007/3-540-47969-4_44}, booktitle={Computer Vision — ECCV 2002}, publisher={Springer Berlin Heidelberg}, author={Pérez, P. and Hue, C. and Vermaak, J. and Gangnet, M.}, year={2002}, pages={661–675} }
Metadata:Crossref

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF