Back

The Recognition of Human Movement Using Temporal Templates

A. BobickJames W. Davis

article2001TPAMI3,138 citations

Introduces temporal templates—combining Motion-Energy and Motion-History Images—to capture the spatial layout and recency of movement for efficient, real-time human action recognition without requiring 3D body reconstruction.

Listen

The paper introduces a view-based method for representing and recognizing human movements directly from image sequences, without first reconstructing three-dimensional body models. This approach addresses the growing need in computer vision to label actions in scenes for applications such as wireless interfaces and interactive environments, where traditional geometry-focused techniques fall short for consistent, view-dependent motions.

The work set out to demonstrate that a compact, two-component temporal template—capturing both the spatial extent of motion and its recency—could discriminate among a range of whole-body movements when matched against stored view-specific examples. The method constructs a binary motion-energy image to show where motion occurred and a motion-history image whose pixel intensity encodes how recently motion took place at each location. These templates are summarized by translation- and scale-invariant Hu moments, then compared via a Mahalanobis distance; a backward-looking search over integration windows enables automatic temporal segmentation and linear speed invariance.

Experiments recorded seven views of 18 aerobics exercises performed by an instructor, with a second subject providing test sequences captured at roughly 30° and 60° offsets. A single-camera match correctly identified 12 of the 18 movements, and the distance gap between the nearest incorrect match and the true class was typically small relative to the median distance across the set. Adding a second orthogonal camera reduced errors to three. A separate real-time implementation, trained on four people performing sitting, waving, and crouching from 45° view increments, ran at 9 Hz on standard hardware and produced few misclassifications when background subtraction replaced simple differencing.

These results indicate that coarse global shape statistics from the paired templates supply sufficient discrimination for many repetitive or articulated actions, while remaining computationally light enough for live deployment. The approach therefore supports low-cost recognition pipelines in constrained settings such as interactive rooms or monitoring stations, where full three-dimensional recovery is unnecessary and often impractical.

Further work should focus on isolating relevant motion when multiple people or camera movement are present, perhaps by combining overhead tracking with body-centered compensation. Extending the representation to automatically mask variable body parts (for example, legs during throwing) would broaden applicability. The current evidence is strongest for isolated, well-lit actions performed against static backgrounds; performance in cluttered or multi-person scenes remains untested at scale and would benefit from additional subjects and more robust motion detection before operational use.

Georgia Institute of TechnologyThe Ohio State Universityimage sequenceviewpointtemporal correlationsTemporal TemplateMotion‑Energy ImageMotion‑History Image
Bobick et al (2001).pdf
  • Paper: Statistical Pattern Recognition: A Review, Anil K. Jain et al. (2000). Reviewing foundational statistical pattern recognition principles provides essential context for the classification and decision-theoretic tools used in temporal template matching.
  • Paper: Lucas-Kanade 20 Years On: A Unifying Framework, Simon Baker et al. (2004). Understanding advanced image alignment and tracking frameworks clarifies the motion estimation and temporal registration steps required for template-based video analysis.
Cover for The Recognition of Human Movement Using Temporal Templates

Abstract

A new view-based approach to the representation and recognition of human movement is presented. The basis of the representation is a temporal template—a static vector-image where the vector value at each point is a function of the motion properties at the corresponding spatial location in an image sequence. Using aerobics exercises as a test domain, we explore the representational power of a simple, two component version of the templates: The first value is a binary value indicating the presence of motion and the second value is a function of the recency of motion in a sequence. We then develop a recognition method matching temporal templates against stored instances of views of known actions. The method automatically performs temporal segmentation, is invariant to linear changes in speed, and runs in real-time on standard platforms.

Topics

image sequenceviewpointtemporal correlationscharacter recognitionrecognition algorithm

Keywords

Temporal TemplateMotion‑Energy ImageMotion‑History ImageHu MomentMahalanobis DistanceAerobic ExerciseView‑InvariantReal‑TimeOptical FlowSegmentation

Table of Contents

  • 2 PRIOR WORK
  • 2.1 Generic Human Model Recovery
  • 2.1.1 Three-Dimensional Movement Recognition
  • 2.2 Appearance-Based Models
  • 2.3 Motion-Based Recognition
  • 3 TEMPORAL TEMPLATES
  • 3.1 Motion-Energy Images
  • 3.2 Motion-History Images
  • 3.3 Extending Temporal Templates
  • 4 DISCRIMINATION
  • 4.1 Matching Temporal Templates
  • 4.2 Testing on Aerobics Data: One Camera
  • 4.3 Combining Multiple Views
  • 5 SEGMENTATION AND RECOGNITION
  • 6 EXTENSIONS, PROBLEMS, AND APPLICATIONS
  • 6.1 Handling Incidental Motion
  • 6.2 The KidsRoom: An Application
  • APPENDIX
  • IMAGE MOMENTS
  • ACKNOWLEDGMENTS
  • REFERENCES

Knowls

  1. Knowl 1 — Motion-Energy Image (MEI) and Motion-History Image (MHI) Temporal Templates

    model/method

    A temporal template is a 2D view-specific vector-image that represents human movement directly from image motion without requiring 3D body reconstruction. The representation decomposes motion into spatial extent (where motion occurs) and temporal dynamics (how motion develops over time) across an observation window of duration τ\tau.

    Let I(x,y,t)I(x, y, t) be an image sequence and D(x,y,t)∈{0,1}D(x, y, t) \in \{0, 1\} be a binary motion image indicating pixel-level motion (e.g., generated via image differencing or background subtraction). The binary Motion-Energy Image (MEI), denoted Eτ(x,y,t)E_\tau(x, y, t), is defined as the temporal union of motion regions over duration τ\tau:

    Eτ(x,y,t)=⋃i=0τ−1D(x,y,t−i)E_\tau(x, y, t) = \bigcup_{i=0}^{\tau-1} D(x, y, t - i)

    The Motion-History Image (MHI), denoted Hτ(x,y,t)H_\tau(x, y, t), is a scalar-valued image where pixel intensity reflects the recency of motion. It is updated recursively via a replacement and linear decay operator:

    Hτ(x,y,t)={τif D(x,y,t)=1max⁡(0,Hτ(x,y,t−1)−1)otherwiseH_\tau(x, y, t) = \begin{cases} \tau & \text{if } D(x, y, t) = 1 \\ \max(0, H_\tau(x, y, t - 1) - 1) & \text{otherwise} \end{cases}

    Pixels where motion occurred most recently take maximum intensity τ\tau, while older motion decays toward zero. The MEI is directly obtained by thresholding the MHI above zero: Eτ(x,y,t)=1E_\tau(x, y, t) = 1 if Hτ(x,y,t)>0H_\tau(x, y, t) > 0, and 00 otherwise. This recursive formulation requires storing only the previous MHI and the current frame difference, making temporal template accumulation computationally lightweight and memory-efficient.

  2. Knowl 2 — Scale-Invariant Dynamic Temporal Segmentation and Template Matching

    algorithm

    To recognize movements executed at varying speeds without prior knowledge of start and end times, a backward-looking variable-duration search computes and evaluates temporal templates across a span of durations [τmin⁡,τmax⁡][\tau_{\min}, \tau_{\max}] using nn discrete window lengths.

    Input: Binary motion image D(x,y,t)D(x, y, t), previous MHI Hτmax⁡(x,y,t−1)H_{\tau_{\max}}(x, y, t-1), duration bounds [τmin⁡,τmax⁡][\tau_{\min}, \tau_{\max}], number of search windows nn, statistical models (μMEI,ΣMEI,μMHI,ΣMHI)(\boldsymbol{\mu}_{\text{MEI}}, \boldsymbol{\Sigma}_{\text{MEI}}, \boldsymbol{\mu}_{\text{MHI}}, \boldsymbol{\Sigma}_{\text{MHI}}) for each movement/view class, distance threshold θ\theta.
    Output: Best matching movement label and view angle, or no match.
    Step 1: Compute MHI for the maximum duration τmax⁡\tau_{\max}:
    for each pixel (x,y)(x, y):
        if D(x,y,t)==1D(x, y, t) == 1:
            Hτmax⁡(x,y,t)=τmax⁡H_{\tau_{\max}}(x, y, t) = \tau_{\max}
        else:
            Hτmax⁡(x,y,t)=max⁡(0,Hτmax⁡(x,y,t−1)−1)H_{\tau_{\max}}(x, y, t) = \max(0, H_{\tau_{\max}}(x, y, t-1) - 1)
    Step 2: Derive nn candidate templates via thresholding and scaling:
    Δτ=(τmax⁡−τmin⁡)/(n−1)\Delta\tau = (\tau_{\max} - \tau_{\min}) / (n - 1)
    for k=0k = 0 to n−1n - 1:
        τk=τmax⁡−k⋅Δτ\tau_k = \tau_{\max} - k \cdot \Delta\tau
        for each pixel (x,y)(x, y):
            if Hτmax⁡(x,y,t)>k⋅ΔτH_{\tau_{\max}}(x, y, t) > k \cdot \Delta\tau:
                Hτk(x,y,t)=Hτmax⁡(x,y,t)−k⋅ΔτH_{\tau_k}(x, y, t) = H_{\tau_{\max}}(x, y, t) - k \cdot \Delta\tau
            else:
                Hτk(x,y,t)=0H_{\tau_k}(x, y, t) = 0
            Hˉτk(x,y,t)=Hτk(x,y,t)/τk\bar{H}_{\tau_k}(x, y, t) = H_{\tau_k}(x, y, t) / \tau_k
            if Hˉτk(x,y,t)>0\bar{H}_{\tau_k}(x, y, t) > 0:
                Eτk(x,y,t)=1E_{\tau_k}(x, y, t) = 1
            else:
                Eτk(x,y,t)=0E_{\tau_k}(x, y, t) = 0
    Step 3: Extract invariant moments and match:
    for each candidate duration τk\tau_k:
        Extract 7 Hu moments mMEI\mathbf{m}_{\text{MEI}} from EτkE_{\tau_k} and mMHI\mathbf{m}_{\text{MHI}} from Hˉτk\bar{H}_{\tau_k}
        for each known action/view model jj:
            Calculate Mahalanobis distance dMEI=(mMEI−μMEI,j)TΣMEI,j−1(mMEI−μMEI,j)d_{\text{MEI}} = \sqrt{(\mathbf{m}_{\text{MEI}} - \boldsymbol{\mu}_{\text{MEI}, j})^T \boldsymbol{\Sigma}_{\text{MEI}, j}^{-1} (\mathbf{m}_{\text{MEI}} - \boldsymbol{\mu}_{\text{MEI}, j})}
            if dMEI<θd_{\text{MEI}} < \theta:
                Calculate dMHI=(mMHI−μMHI,j)TΣMHI,j−1(mMHI−μMHI,j)d_{\text{MHI}} = \sqrt{(\mathbf{m}_{\text{MHI}} - \boldsymbol{\mu}_{\text{MHI}, j})^T \boldsymbol{\Sigma}_{\text{MHI}, j}^{-1} (\mathbf{m}_{\text{MHI}} - \boldsymbol{\mu}_{\text{MHI}, j})}
                Record match jj with combined distance d=dMEI+dMHId = d_{\text{MEI}} + d_{\text{MHI}}
    return movement class corresponding to the smallest distance dd

    Scaling HτkH_{\tau_k} by 1/τk1/\tau_k normalizes intensities to the range [0,1][0, 1], making the moment descriptors invariant to linear changes in movement speed. Hierarchical filtering (testing MEI first, then verifying MHI agreement) minimizes unnecessary covariance evaluations.

  3. Knowl 3 — Moment-Based Invariant Shape Descriptors for Temporal Templates

    equation

    To achieve translation and scale invariance for temporal template matching, statistical descriptions of MEIs and MHIs are computed using 2D geometric moment invariants.

    For a 2D intensity distribution ρ(x,y)\rho(x, y), the (p+q)(p+q)th-order raw moment mpqm_{pq} and central moment μpq\mu_{pq} are defined as:

    mpq=∫−∞∞∫−∞∞xpyqρ(x,y) dx dym_{pq} = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} x^p y^q \rho(x, y) \, dx \, dy

    μpq=∫−∞∞∫−∞∞(x−xˉ)p(y−yˉ)qρ(x,y) dx dy\mu_{pq} = \int_{-\infty}^{\infty} \int_{-\infty}^{\infty} (x - \bar{x})^p (y - \bar{y})^q \rho(x, y) \, dx \, dy

    where xˉ=m10/m00\bar{x} = m_{10}/m_{00} and yˉ=m01/m00\bar{y} = m_{01}/m_{00}. Scale invariance is achieved by normalizing the central moments:

    ηpq=μpqμ00γ,where γ=p+q2+1(p+q≥2)\eta_{pq} = \frac{\mu_{pq}}{\mu_{00}^\gamma}, \quad \text{where } \gamma = \frac{p+q}{2} + 1 \quad (p+q \ge 2)

    Seven invariant Hu moments ϕ1,…,ϕ7\phi_1, \dots, \phi_7 are computed from the normalized central moments up to order 3:

    ϕ1=η20+η02\phi_1 = \eta_{20} + \eta_{02}

    ϕ2=(η20−η02)2+4η112\phi_2 = (\eta_{20} - \eta_{02})^2 + 4\eta_{11}^2

    ϕ3=(η30−3η12)2+(3η21−η03)2\phi_3 = (\eta_{30} - 3\eta_{12})^2 + (3\eta_{21} - \eta_{03})^2

    ϕ4=(η30+η12)2+(η21+η03)2\phi_4 = (\eta_{30} + \eta_{12})^2 + (\eta_{21} + \eta_{03})^2

    ϕ5=(η30−3η12)(η30+η12)[(η30+η12)2−3(η21+η03)2]+(3η21−η03)(η21+η03)[3(η30+η12)2−(η21+η03)2]\phi_5 = (\eta_{30} - 3\eta_{12})(\eta_{30} + \eta_{12})[(\eta_{30} + \eta_{12})^2 - 3(\eta_{21} + \eta_{03})^2] + (3\eta_{21} - \eta_{03})(\eta_{21} + \eta_{03})[3(\eta_{30} + \eta_{12})^2 - (\eta_{21} + \eta_{03})^2]

    ϕ6=(η20−η02)[(η30+η12)2−(η21+η03)2]+4η11(η30+η12)(η21+η03)\phi_6 = (\eta_{20} - \eta_{02})[(\eta_{30} + \eta_{12})^2 - (\eta_{21} + \eta_{03})^2] + 4\eta_{11}(\eta_{30} + \eta_{12})(\eta_{21} + \eta_{03})

    ϕ7=(3η21−η03)(η30+η12)[(η30+η12)2−3(η21+η03)2]−(η30−3η12)(η21+η03)[3(η30+η12)2−(η21+η03)2]\phi_7 = (3\eta_{21} - \eta_{03})(\eta_{30} + \eta_{12})[(\eta_{30} + \eta_{12})^2 - 3(\eta_{21} + \eta_{03})^2] - (\eta_{30} - 3\eta_{12})(\eta_{21} + \eta_{03})[3(\eta_{30} + \eta_{12})^2 - (\eta_{21} + \eta_{03})^2]

    A movement view is characterized by a 14-dimensional feature vector consisting of the 7 Hu moments of its MEI and the 7 Hu moments of its MHI, modeled by mean vectors μ\boldsymbol{\mu} and covariance matrices Σ\boldsymbol{\Sigma} across training instances.

  4. Knowl 4 — Multi-Camera Fusion via Joint Mahalanobis Distance Minimization

    model/method

    Because single-view temporal templates can be ambiguous when different actions produce similar 2D projections (e.g., sitting versus crouching viewed frontally), multiple camera views are combined to improve movement discrimination.

    When camera angular calibration is known (e.g., two orthogonal cameras with a relative angular separation of Δθ=90∘\Delta\theta = 90^\circ), recognition selects the movement class kk and viewpoint θ\theta that minimizes the joint Mahalanobis distance across both views:

    k∗=arg⁡min⁡kmin⁡θ[dM(m1,modelk(θ))+dM(m2,modelk(θ+Δθ))]k^* = \arg\min_k \min_\theta \left[ d_M\left(\mathbf{m}_1, \text{model}_k(\theta)\right) + d_M\left(\mathbf{m}_2, \text{model}_k(\theta + \Delta\theta)\right) \right]

    where m1\mathbf{m}_1 and m2\mathbf{m}_2 are the moment feature vectors from camera 1 and camera 2, and dM(x,model)=(x−μ)TΣ−1(x−μ)d_M(\mathbf{x}, \text{model}) = \sqrt{(\mathbf{x} - \boldsymbol{\mu})^T \boldsymbol{\Sigma}^{-1} (\mathbf{x} - \boldsymbol{\mu})}.

    If the inter-camera angular relationship is unknown, labeling consistency is enforced by requiring both cameras to select the same movement label kk, choosing the movement that minimizes the sum of distances over independent viewpoint choices:

    k∗=arg⁡min⁡k[min⁡θ1dM(m1,modelk(θ1))+min⁡θ2dM(m2,modelk(θ2))]k^* = \arg\min_k \left[ \min_{\theta_1} d_M\left(\mathbf{m}_1, \text{model}_k(\theta_1)\right) + \min_{\theta_2} d_M\left(\mathbf{m}_2, \text{model}_k(\theta_2)\right) \right]

    Relative angular ordering (e.g., camera 1 is to the left of camera 2) can further constrain the candidate viewpoint pairs.

  5. Knowl 5 — Aerobics Movement Recognition Performance across Single- and Dual-Camera Setups

    data/table

    The temporal template representation was evaluated on an 18-action aerobic exercise dataset recorded across 7 viewing angles (−90∘-90^\circ to +90∘+90^\circ in 30∘30^\circ increments, yielding 126 view/move models). Test moves were executed by a separate subject and observed by one camera (30∘30^\circ off frontal) or two orthogonal cameras (30∘30^\circ left and 60∘60^\circ right).

    Single Camera (30∘30^\circ Off Frontal) Two Cameras (90∘90^\circ Angular Interval)
    Test Closest Closest Correct Rank Closest Closest Correct Rank
    Move Move Dist Dist (Single) Move Dist Dist (Dual)
    1 4 1.43 1.44 2 1 2.13 2.13 1
    2 2 3.14 3.14 1 2 12.92 12.92 1
    3 3 3.08 3.08 1 3 7.17 7.17 1
    4 4 0.47 0.47 1 4 1.07 1.07 1
    5 5 6.84 6.84 1 5 16.42 16.42 1
    6 10 0.32 0.61 7 6 0.88 0.88 1
    7 7 0.97 0.97 1 7 3.02 3.02 1
    8 8 20.47 20.47 1 8 36.76 36.76 1
    9 8 1.05 1.77 4 8 5.10 6.74 3
    10 10 0.14 0.14 1 10 0.68 0.68 1
    11 11 0.24 0.24 1 11 1.20 1.20 1
    12 12 0.79 0.79 1 12 2.77 2.77 1
    13 6 0.13 0.25 3 13 0.57 0.57 1
    14 14 4.01 4.01 1 14 6.07 6.07 1
    15 15 0.34 0.34 1 15 2.28 2.28 1
    16 15 1.03 1.04 2 15 1.86 2.35 2
    17 17 0.65 0.65 1 8 2.67 3.24 3
    18 10 0.48 0.51 4 18 1.18 1.18 1

    Single-camera matching correctly identifies 12 of 18 moves. In most misclassified cases (moves 1, 9, 13, 16, 18), the distance to the correct move is very close to the minimum distance compared to the median distance across all classes. Combining two orthogonal views improves accuracy to 15 of 18 correct classifications (reducing errors from 6 to 3), resolving viewing ambiguities where motions look similar from one angle but distinct from another.

  6. Knowl 6 — Real-Time Movement Recognition System Setup and Performance

    experimental setup

    A real-time prototype was implemented to demonstrate online segmentation and recognition on standard hardware.

    • Hardware and Sensors: Two CCD cameras interfaced to a Silicon Graphics Indy workstation (200 MHz).
    • Image Resolution: Captured video was downsampled and digitized to 160×120160 \times 120 pixels.
    • Movement Vocabulary: 3 actions (sitting, arm waving, crouching) executed across a 180∘180^\circ view circle.
    • Training Data: 4 human subjects performing each movement, sampled every 45∘45^\circ around the view circle.
    • Search Window Parameters: τmin⁡=11\tau_{\min} = 11 frames (approx. 1 second), τmax⁡=19\tau_{\max} = 19 frames (approx. 2 seconds), and n=6n = 6 integration window steps.
    • Processing Rate: The entire vision pipeline (image differencing, MHI decay/thresholding, Hu moment extraction, and Mahalanobis distance calculation) executed at approximately 9 Hz9\text{ Hz}.

    Because the distance matching step contributes negligible computational overhead, expanding the movement model database does not decrease frame rate, affecting only classification accuracy.

  7. Knowl 7 — Background-Subtracted Temporal Templates for Interactive Story Spaces

    model/method

    In interactive multi-person environments such as The KidsRoom, motion differencing can produce noisy temporal templates when users move non-critical body parts or hold parts of their bodies momentarily stationary during an action (e.g., dancing or spinning).

    To increase robustness against incidental motion variations, temporal templates are generated by accumulating binary background-subtracted body silhouettes rather than motion-differenced frames: D(x,y,t)D(x, y, t) is replaced with the binary foreground mask S(x,y,t)S(x, y, t).

    Under this formulation, stationary foreground body parts continuously refresh the template, preventing gaps from appearing inside the body region when a subject moves slowly. While this modification slightly reduces the fine-grained motion discrimination power, it substantially improves tracking and gesture recognition stability in natural, unconstrained interactive spaces.

  8. Knowl 8 — Failure Modes and Operational Constraints of Temporal Templates

    limitation

    The temporal template approach exhibits specific operational failure modes:

    1. Motion Self-Occlusion and Overwrite: Because the MHI relies on a pixel-wise recency decay operator, any newer motion passing over a previously active spatial location overwrites earlier motion history, destroying past temporal information at those coordinates.
    2. Incidental Body Motion: Unconstrained movements of secondary limbs (e.g., leg position during a ball throw) alter global moment statistics, requiring spatial masking or region isolation to prevent classification failures.
    3. Motion Detection Sensitivity: When using simple frame differencing, slow body movement or low-contrast clothing against the background causes failed motion detection, resulting in fragmented MHIs.
    4. Camera Motion and Locomotion: Global egomotion or whole-body locomotion corrupts the temporal template by mixing subject trajectory with internal limb motion, requiring body-centered motion field compensation.
    5. Multi-Person Occlusion: When multiple individuals overlap in the camera view, bounding-box motion isolation fails, requiring multi-camera or overhead tracking configurations to isolate clear views.

Coverage note — None was omitted; all key contributions including temporal template representations, moment-based matching, dynamic temporal segmentation, multiview fusion, empirical aerobics evaluations, real-time implementation, and application adaptations have been captured.

References

  1. 1.J. Aggarwal and N. Nandhakumar, "On the Computation of Motion of Sequences of Images—A Review," Proc. IEEE, vol. 69, no. 5, pp. 917-934, 1988.
  2. 2.J. Aggarwal and Q. Cai, "Human Motion Analysis: A Review," Computer Vision and Image Understanding, vol. 73, no. 3, pp. 428-440, 1999.
  3. 3.K. Akita, "Image Sequence Analysis of Real World Human Motion," Pattern Recognition, vol. 17, no. 1, pp. 73-83, 1984.
  4. 4.M.J. Black and Y. Yacoob, "Tracking and Recognizing Rigid and Non-Rigid Facial Motions Using Local Parametric Model of Image Motion," Proc. Int'l Conf. Computer Vision, pp. 374-381, 1995.
  5. 5.A.F. Bobick, S.S. Intille, J.W. Davis, F. Baird, L.W. Campbell, Y. Ivanov, C.S. Pinhanez, A. Schütte, and A. Wilson, "The KidsRoom: A Perceptually-Based Interactive and Immersive Story Environment," Presence, vol. 8, no. 4, pp. 368-393, Aug. 1999.
  6. 6.A. Bobick, "Movement, Activity, and Action: The Role of Knowledge in the Perception of Motion," Philosophical Trans. Royal Soc. London, vol. 352, pp. 1257-1265, 1997.
  7. 7.A. Bobick and J. Davis, "An Appearance-Based Representation of Action," Proc. Int'l Conf. Pattern Recognition, pp. 307-312, 1996.
  8. 8.L. Campbell and A. Bobick, "Recognition of Human Body Motion Using Phase Space Constraints," Proc. Int'l Conf. Computer Vision, pp. 624-630, 1995.
  9. 9.Y. Cui, D. Swets, and J. Weng, "Learning-Based Hand Sign Recognition Using Shoslif-m," Proc. Int'l Conf. Computer Vision, pp. 631-636, 1995.
  10. 10.T. Darrell and A. Pentland, "Space-Time Gestures," Proc. Computer Vision and Pattern Recognition, pp. 335-340, 1993.
  11. 11.J. Davis and A. Bobick, "The Representation and Recognition of Human Movement Using Temporal Templates," Proc. Computer Vision and Pattern Recognition, pp. 928-934, 1997.
  12. 12.I. Essa and A. Pentland, "Coding, Analysis, Interpretation, and Recognition of Facial Expressions," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp. 757-763, July 1997.
  13. 13.W. Freeman and M. Roth, "Orientation Histogram for Hand Gesture Recognition," Proc. Int'l Workshop Automatic Face and Gesture Recognition, pp. 296-301, 1995.
  14. 14.D.M. Gavrila and L.S. Davis, "3D Model-Based Tracking of Humans in Action: A Multiview Approach," Proc. Computer Vision and Pattern Recognition, pp. 73-80, 1996.
  15. 15.L. Goncalves, E. DiBernardo, E. Ursella, and P. Perona, "Monocular Tracking of the Human Arm in 3D," Proc. Int'l Conf. Computer Vision, pp. 764-770, Aug. 1995.
  16. 16.D. Hogg, "Model-Based Vision: A Paradigm to See a Walking Person," Image and Vision Computing, vol. 1, no. 1, pp. 5-20, 1983.
  17. 17.M. Hu, "Visual Pattern Recognition by Moment Invariants," IRE Trans. Information Theory, vol. 8, no. 2, pp. 179-187, 1962.
  18. 18.D. Jones and J. Malik, "Computational Framework for Determining Stereo Correspondence from a Set of Linear Spatial Filters," Image and Vision Computing, vol. 10, no. 10, pp. 699-708, 1992.
  19. 19.S. Ju, M. Black, and Y. Yacoob, "Cardboard People: A Parameterized Model of Articulated image Motion," Proc. Second Int'l Conf. Automatic Face and Gesture Recognition, pp. 38-44, Oct. 1996.
  20. 20.J. Little and J. Boyd, "Describing Motion for Recognition," Int'l Symp. Computer Vision, pp. 235-240, Nov. 1995.
  21. 21.P. Maes, T. Darrell, B. Blumberg, and A. Pentland, "The ALIVE System: Wireless, Full-Body Interaction with Autonomous Agents," ACM Multimedia Systems, 1996.
  22. 22.R. Polana and R. Nelson, "Low Level Recognition of Human Motion," Proc. IEEE Workshop Non-Rigid and Articulated Motion, pp. 77-82, 1994.
  23. 23.J. Rehg and T. Kanade, "Model-Based Tracking of Self-Occluding Articulated Objects," Proc. Int'l Conf. Computer Vision, pp. 612-617, 1995.
  24. 24.K. Rohr, "Towards Model-Based Recognition of Human Movements in Image Sequences," CVGIP, Image Understanding, vol. 59, no. 1, pp. 94-115, 1994.
  25. 25.E. Shavit and A. Jepson, "Motion Understanding Using Phase Portraits," Proc. IJCAI Workshop: Looking at People, 1993.
  26. 26.J.M. Siskind, "Grounding Language in Perception," Artificial Intelligence Rev., vol. 8, pp. 371-391, 1995.
  27. 27.A. Wilson and A. Bobick, "Learning Visual Behavior for Gesture Analysis," Proc. IEEE Int'l. Symp. Computer Vision, Nov. 1995.
  28. 28.Y. Yacoob and L. Davis, "Recognizing Human Facial Expressions Form Long Image Sequences Using Optical Flow," IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 18, pp. 636-642, 1996.
  29. 29.J. Yamato, J. Ohya, and K. Ishii, "Recognizing Human Action in Time Sequential Images Using Hidden Markov Models," Proc. Computer Vision and Pattern Recognition, pp. 379-385, 1992.

Citation

MLA
Bobick, A. F., and J. W. Davis. “The Recognition of Human Movement Using Temporal Templates”. IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 23, no. 3, 2001, pp. 257–67, https://doi.org/10.1109/34.910878.
APA
Bobick, A. F., & Davis, J. W. (2001). The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(3), 257–267. https://doi.org/10.1109/34.910878
Chicago
Bobick, A. F., and J. W. Davis. 2001. “The Recognition of Human Movement Using Temporal Templates”. IEEE Transactions on Pattern Analysis and Machine Intelligence 23 (3): 257–67. https://doi.org/10.1109/34.910878.
Harvard
Bobick, A.F. and Davis, J.W. (2001) “The recognition of human movement using temporal templates”, IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(3), pp. 257–267. Available at: https://doi.org/10.1109/34.910878.
Vancouver
1. Bobick AF, Davis JW (2001) The recognition of human movement using temporal templates. IEEE Transactions on Pattern Analysis and Machine Intelligence 23:257–267

BibTeX

@article{Bobick_2001, title={The recognition of human movement using temporal templates}, volume={23}, ISSN={0162-8828}, url={http://dx.doi.org/10.1109/34.910878}, DOI={10.1109/34.910878}, number={3}, journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, publisher={Institute of Electrical and Electronics Engineers (IEEE)}, author={Bobick, A.F. and Davis, J.W.}, year={2001}, month=Mar, pages={257–267} }
Metadata:Crossref

Access the Paper

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

Open PDF