Back

Human Detection Using Oriented Histograms of Flow and Appearance

Navneet DalalBill TriggsCordelia Schmid

article2006ECCV1,965 citations

Combines oriented histograms of differential optical flow with Histogram of Oriented Gradient appearance descriptors to achieve a tenfold reduction in false alarms for video-based human detection in complex dynamic scenes.

Listen

Reliable automated human detection in video is essential for technologies such as automotive pedestrian safety systems, surveillance, and automated film analysis. While appearance-based visual detectors have advanced, they struggle with high false alarm rates in complex real-world scenes. Prior motion-assisted detection systems typically assume fixed cameras and stationary backgrounds, making them ineffective in dynamic environments where camera pan, tilt, or vehicle movement introduce substantial visual noise. The article addresses this operational gap by developing and evaluating motion-based feature schemes that isolate characteristic human movement even when both the camera and background are in motion.

The investigation set out to demonstrate whether integrating differential optical flow—which captures patterns of apparent pixel motion between consecutive video frames—with established static appearance descriptors could significantly reduce false alarms without sacrificing detection accuracy. The authors designed and evaluated several feature representations, focusing on motion boundary histograms and internal motion histograms based on differences across adjacent grid cells. These descriptors were paired with static Histogram of Oriented Gradient visual descriptors and classified using linear Support Vector Machines, which are practical, fast, and scalable learning algorithms. To support rigorous evaluation, the system was trained and tested on challenging real-world footage containing over 4,400 human annotations derived from multiple feature films and video sequences featuring complex lighting, varied clothing, diverse poses, and camera motion.

The findings show that combining differential optical flow with static appearance descriptors reduces the false alarm rate by a factor of 10 relative to the best appearance-only methods. For instance, in standard video testing, the combined detector achieved a false positive rate of only 1 in 20,000 scanned windows at an 8% miss rate. Interestingly, internal motion histograms that measure motion differences across neighboring cells outperformed motion-boundary schemes when combined with appearance data, because they effectively capture complementary limb dynamics rather than duplicating static edge information. Furthermore, a fast, unregularized optical flow algorithm computed in roughly one second per frame outperformed slower, heavily smoothed flow methods by a factor of three in reducing false alarms, as it preserved sharp, fine-grained limb movements. Finally, testing on entirely static scenes confirmed that incorporating motion features does not degrade performance when motion is absent.

These results provide a clear operational pathway for developing real-time, low-latency detection systems in mobile settings like automotive collision avoidance and dynamic surveillance. The substantial reduction in false alarms directly translates to lower operational risk, fewer unnecessary automated interventions, and improved system reliability. The analysis also revealed that modular classification architectures, such as a mixture of experts that evaluates appearance and motion separately before merging scores, offer slight performance gains and lower computational memory requirements during training compared to monolithic models.

For practical implementation, engineering teams should adopt internal motion histograms paired with lightweight, unregularized multi-scale optical flow rather than computationally heavy, over-smoothed motion estimators. Future development should focus on integrating multi-stage rejection cascades to further accelerate processing speed, refining part-based models to handle severe occlusions, and enforcing temporal tracking across multiple video frames rather than treating consecutive frame pairs independently. While the findings provide high confidence for upright and largely visible individuals across diverse environments, cautious validation is recommended before deploying the system in scenarios characterized by heavy body occlusions or non-upright body postures.

INRIAbackground cluttercoding schemedescriptorshuman detectionpedestrian detectionHOG descriptors
Dalal et al (2006).pdf
  • Paper: Action recognition by dense trajectories, Heng Wang et al. (2011). Building on motion-boundary histograms for detection, this later work turns dense optical-flow trajectories and motion descriptors toward action recognition in challenging video.
  • Paper: Future Frame Prediction for Anomaly Detection - A New Baseline, Wen Liu et al. (2017). This later anomaly-detection method extends motion cues into future-frame prediction, using optical flow as a temporal constraint for recognizing unexpected events.
Cover for Human Detection Using Oriented Histograms of Flow and Appearance

Abstract

Detecting humans in films and videos is a challenging problem owing to the motion of the subjects, the camera and the background and to variations in pose, appearance, clothing, illumination and background clutter. We develop a detector for standing and moving people in videos with possibly moving cameras and backgrounds, testing several different motion coding schemes and showing empirically that orientated histograms of differential optical flow give the best overall performance. These motion-based descriptors are combined with our Histogram of Oriented Gradient appearance descriptors. The resulting detector is tested on several databases including a challenging test set taken from feature films and containing wide ranges of pose, motion and background variations, including moving cameras and backgrounds. We validate our results on two challenging test sets containing more than 4400 human examples. The combined detector reduces the false alarm rate by a factor of 10 relative to the best appearance-based detector, for example giving false alarm rates of 1 per 20,000 windows tested at 8% miss rate on our Test Set 1.

Topics

background cluttercoding schemedescriptorsobject appearanceoptical flowstest samplescolor histogramdetection algorithmclutter backgroundpose variation

Keywords

human detectionpedestrian detectionHOG descriptorsmotion descriptorsoptical flow methodsappearance descriptorsorientation histogramsdifferential flowspatial gradients

Table of Contents

  • 1 Introduction
  • 2 Previous Work
  • 3 Overall Architecture
  • 4 Appearance Descriptors
  • 5 Motion Descriptors
  • 5.1 Motion Boundary Based Coding
  • 5.2 Internal / Relative Dynamics Based Coding
  • 5.3 Descriptor Parameters
  • 6 Optical Flow Estimation
  • 7 Data Sets
  • 8 Experiments
  • 9 Summary and Conclusions
  • References

Knowls

  1. Knowl 1 — Combined Optical Flow and Appearance Person Detector Architecture

    model/method

    The human detection framework combines static appearance descriptors with motion descriptors computed from consecutive video frames to detect upright, fully visible standing and moving humans under arbitrary camera and background motion.

    At detection time, a detection window of size 64×12864 \times 128 pixels is scanned across the image at multiple positions and scales. For each window, visual appearance features (Histogram of Oriented Gradients, or HOG) and motion features (oriented histograms of differential optical flow) are extracted and concatenated into a monolithic descriptor vector, or evaluated through a two-stage classifier. A linear Support Vector Machine (SVM) evaluates each window descriptor to output a detection score.

    To identify bounding boxes across scale and spatial coordinates, negative classifier scores are set to zero, and a 3D position-scale mean-shift mode-seeking process is applied to group overlapping detection windows and locate local score maxima. Peaks exceeding a preset detection threshold are returned as detections without enforcing inter-frame temporal tracking.

    Training proceeds in two stages:

    1. An initial linear SVM is trained on manually labeled positive bounding windows (including their horizontal mirror reflections) and an initial set of randomly sampled negative bounding boxes from background images.
    2. The initial detector exhaustively evaluates negative training images to collect hard false positives ("hard negatives"). As many hard negative feature vectors as fit within memory limits (e.g., 1.5 GB RAM) are added to the training set to train the final linear SVM.

  2. Knowl 2 — Motion Boundary Histograms for Differential Flow Coding

    model/method

    Motion Boundary Histograms (MBH) encode human motion boundaries while suppressing smooth optical flow fields caused by camera translation, pan, tilt, and roll. Because smooth background and camera motions produce locally uniform flow across depth boundaries, taking spatial derivatives of the optical flow cancels out uniform camera motion and highlights motion discontinuities (such as moving limbs and body contours).

    Let the estimated optical flow field between two consecutive frames be denoted by w(x,y)=(Ix(x,y),Iy(x,y))\mathbf{w}(x, y) = (I^x(x, y), I^y(x, y)), where IxI^x and IyI^y represent the horizontal and vertical velocity components. The MBH descriptor treats IxI^x and IyI^y as two independent scalar images:

    1. Spatial derivatives are computed separately for each flow component using centered 1D difference masks [−1,0,1][-1, 0, 1]: Ixx=∂Ix∂x,Iyx=∂Ix∂y,Ixy=∂Iy∂x,Iyy=∂Iy∂yI^x_x = \frac{\partial I^x}{\partial x}, \quad I^x_y = \frac{\partial I^x}{\partial y}, \quad I^y_x = \frac{\partial I^y}{\partial x}, \quad I^y_y = \frac{\partial I^y}{\partial y}
    2. For each flow channel, local gradient orientations θx=arctan⁡(Iyx/Ixx)\theta^x = \arctan(I^x_y / I^x_x) and θy=arctan⁡(Iyy/Ixy)\theta^y = \arctan(I^y_y / I^y_x) and gradient magnitudes mx=(Ixx)2+(Iyx)2m^x = \sqrt{(I^x_x)^2 + (I^x_y)^2} and my=(Ixy)2+(Iyy)2m^y = \sqrt{(I^y_x)^2 + (I^y_y)^2} are determined at each pixel.
    3. Each pixel votes into orientation histogram bins (e.g., 9 orientation bins over 0∘0^\circ--180∘180^\circ or 0∘0^\circ--360∘360^\circ) weighted by its gradient magnitude mxm^x or mym^y, spatially accumulated over cells (e.g., 8×88 \times 8 pixels).
    4. Histograms are normalized across overlapping spatial blocks of cells (e.g., 2×22 \times 2 cells) using hysteresis-thresholded L2L_2 normalization.

    Keeping separate orientation histograms for the horizontal (IxI^x) and vertical (IyI^y) flow components yields greater discriminative performance than combining their gradients via winner-take-all voting.

  3. Knowl 3 — Internal Motion Histograms for Relative Dynamics and Limb Motions

    model/method

    Internal Motion Histograms (IMH) are differential motion descriptors designed to capture relative limb displacements and internal dynamics within a human silhouette, providing motion cues that complement static boundary appearance.

    Rather than voting based on the spatial orientation of flow gradients as in Motion Boundary Histograms, IMH computes flow difference vectors between spatial positions and votes based on the orientation of the relative flow vector itself:

    • IMHdiff (Differential IMH): Computes fine-scale spatial flow derivatives (Ixx,Ixy)(I^x_x, I^y_x) and (Iyx,Iyy)(I^x_y, I^y_y). Histograms are constructed by binning the vector orientation arctan⁡(Ixy/Ixx)\arctan(I^y_x / I^x_x) weighted by (Ixx)2+(Ixy)2\sqrt{(I^x_x)^2 + (I^y_x)^2}, and similarly for the yy-derivative components.

    • IMHcd (Center-Difference IMH): Tiled using 3×33 \times 3 blocks of spatial cells (where each cell matches approximate limb width, such as 8×88 \times 8 pixels). For each of the 8 outer cells in the block, pixel-wise flow differences are computed relative to the corresponding pixel in the central cell: Δwi(u,v)=wi(u,v)−wcenter(u,v)for i∈{1,…,8}\Delta \mathbf{w}_i(u, v) = \mathbf{w}_i(u, v) - \mathbf{w}_{\text{center}}(u, v) \quad \text{for } i \in \{1, \dots, 8\} The orientation of Δwi\Delta \mathbf{w}_i votes into a 6- or 9-bin orientation histogram weighted by ∥Δwi∥\|\Delta \mathbf{w}_i\|. The resulting 8 cell histograms are normalized jointly as a single block descriptor.

    • IMHmd (Mean-Difference IMH): Computes relative flow differences for each pixel relative to the average flow of the corresponding pixels across all 9 cells in the 3×33 \times 3 block, forming 9 normalized cell histograms per block.

    • IMHwd (Wavelet-Difference IMH): Applies 3×33 \times 3 Haar-wavelet-like difference operators across block cells to encode directional spatial flow differences.

  4. Knowl 4 — Fast Damped Local Least-Squares Optical Flow Estimation for Detection

    algorithm

    For human detection using differential flow histograms, avoiding spatial regularization/smoothing is essential: explicitly regularized smooth flow blurs motion boundaries and internal limb dynamics, degrading combined detection performance. The optical flow w=(Ix,Iy)T\mathbf{w} = (I^x, I^y)^T is estimated using an unregularized, multiscale damped local linear least squares formulation over small N×NN \times N pixel neighborhoods.

    Input: Consecutive image frames ItI_t and It+1I_{t+1}, neighborhood window size N=5N = 5, damping parameter β>0\beta > 0, pyramid downsampling factor s=1.3s = 1.3, number of scales KK
    Output: Dense optical flow field w(x,y)=(Ix(x,y),Iy(x,y))T\mathbf{w}(x, y) = (I^x(x, y), I^y(x, y))^T
    Construct multiscale Gaussian pyramids for ItI_t and It+1I_{t+1} with scale factor ss across levels k=K,…,1k = K, \dots, 1
    Initialize flow field wK(x,y)←(0,0)T\mathbf{w}_K(x, y) \leftarrow (0, 0)^T at the coarsest level KK
    for level k=Kk = K down to 1 do
        if level k<Kk < K then
            Upsample and scale flow field wk+1\mathbf{w}_{k+1} to initialize wk\mathbf{w}_k
        Warp frame It+1(k)I_{t+1}^{(k)} toward It(k)I_t^{(k)} using the current flow wk\mathbf{w}_k
        Compute spatial gradients Ix,IyI_x, I_y from It(k)I_t^{(k)} and temporal differences It(k)−It+1,warped(k)I_t^{(k)} - I_{t+1,\text{warped}}^{(k)}
        for each pixel (x,y)(x, y) do
            Construct N2×2N^2 \times 2 spatial gradient matrix A=[Ix,Iy]A = [I_x, I_y] over the N×NN \times N patch centered at (x,y)(x, y)
            Construct N2×1N^2 \times 1 temporal difference vector bb over the same patch
            Solve for local flow update: Δw(x,y)←(ATA+βI)−1ATb\Delta \mathbf{w}(x, y) \leftarrow (A^T A + \beta I)^{-1} A^T b
            Update flow: wk(x,y)←wk(x,y)+Δw(x,y)\mathbf{w}_k(x, y) \leftarrow \mathbf{w}_k(x, y) + \Delta \mathbf{w}(x, y)
    return Dense flow field w1\mathbf{w}_1 at the finest image scale

    This unregularized local least-squares flow runs in 1 second on 752×396752 \times 396 video frames and reduces false positives in the combined detector by more than a factor of 3 at an 8% miss rate compared to heavily regularized optical flow algorithms (such as multi-scale nonlinear diffusion).

  5. Knowl 5 — Two-Stage Mixture of Experts Fusion for Appearance and Motion Descriptors

    model/method

    As an alternative to a monolithic classifier trained on the concatenated feature vector of appearance (HOG) and motion (IMH/MBH) descriptors, a two-stage Mixture of Experts architecture trains appearance and motion classifiers independently and combines their scalar outputs.

    1. First Stage: Two independent linear Support Vector Machines (SVMs) are trained separately:
      • An appearance detector trained on static HOG descriptors, yielding real-valued scalar decision score sapp(x)s_{\text{app}}(\mathbf{x}).
      • A motion detector trained on differential optical flow descriptors (e.g., IMHcd), yielding real-valued scalar decision score smot(x)s_{\text{mot}}(\mathbf{x}).
    2. Second Stage: A second-stage linear SVM is trained over the 2D feature space defined by the scalar pair (sapp,smot)(s_{\text{app}}, s_{\text{mot}}).

    Because the second-stage feature space is two-dimensional, millions of hard negative examples can be incorporated during retraining without encountering memory limits. The final decision boundary remains a linear combination of the underlying feature vectors, mitigating training losses from separate feature evaluation and achieving a small performance gain (<1% miss rate improvement) over the monolithic concatenated classifier while offering flexibility for cascade rejection.

  6. Knowl 6 — Empirical Performance Gains of Combining Flow and Appearance Descriptors

    empirical result

    Combining Histogram of Oriented Gradients (HOG) appearance descriptors with Center-Difference Internal Motion Histograms (IMHcd) based on unregularized optical flow reduces the false alarm rate by a factor of 10 relative to the best static HOG detector at equivalent miss rates across video test sets containing camera and background motion (e.g., achieving a false alarm rate of 1 per 20,000 tested windows at an 8% miss rate on Test Set 1).

    Key empirical findings across motion representations:

    • When evaluated as motion-only detectors, Motion Boundary Histograms (MBH) and fine-scale IMHdiff with regularized diffusion optical flow achieve the lowest error rates, outperforming static Haar wavelets.
    • When combined with HOG appearance descriptors, IMHcd and IMHmd paired with unregularized multiscale least-squares optical flow achieve the best overall detection accuracy. Motion boundary features (MBH) and regularized flow exhibit high redundancy with static boundary appearance, whereas unregularized IMHcd captures complementary internal limb dynamics.

  7. Knowl 7 — Detection Performance of Combined Detectors on Purely Static Images

    data/table

    Detectors trained on combined static and video datasets (where flow fields on static training images are set identically to zero) maintain their detection performance when evaluated on purely static test images containing no motion.

    Detector Miss Rate at FPPW 10−310^{-3} Miss Rate at FPPW 10−410^{-4} Miss Rate at FPPW 10−510^{-5}
    SHOG (Static HOG) 6.2% 11.4% 19.8%
    SHOG + IMHcd 5.8% 11.0% 19.8%
    SHOG + ST Diff 5.7% 10.5% 19.7%

    The miss rate is measured against False Positives Per Window (FPPW) on the INRIA Static Test Set (1132 images). Despite the complete absence of optical flow (zero flow vector), the combined detectors (SHOG + IMHcd and SHOG + ST Diff) achieve slightly lower miss rates than the static-only HOG detector at 10−310^{-3} and 10−410^{-4} FPPW and identical performance at 10−510^{-5} FPPW, demonstrating that the motion features do not degrade detection when motion cues are absent.

  8. Knowl 8 — Spatiotemporal Intensity Differencing Descriptor (ST Diff)

    model/method

    Spatiotemporal Differencing (ST Diff) is a baseline motion feature that bypasses optical flow estimation by computing direct intensity differences across consecutive video frames.

    For each pixel (x,y)(x, y) in the frame at time tt, image intensity differences are computed against a 3×33 \times 3 grid of pixels in the frame at time t+1t+1 sampled with a spatial stride of 8 pixels around (x,y)(x, y): ΔIj(x,y)=∣It(x,y)−It+1(x+8⋅uj,y+8⋅vj)∣,uj,vj∈{−1,0,1}\Delta I_j(x, y) = |I_t(x, y) - I_{t+1}(x + 8 \cdot u_j, y + 8 \cdot v_j)|, \quad u_j, v_j \in \{-1, 0, 1\} The absolute differences ΔIj\Delta I_j for each of the 9 displacement offsets j∈{1,…,9}j \in \{1, \dots, 9\} are accumulated over all pixels within an 8×88 \times 8 cell, yielding a 9-bin histogram for each cell. The resulting cell histograms undergo standard overlapping block normalization before classification.

  9. Knowl 9 — Optimal Motion Descriptor Parameters and Flow Precision Requirements

    experimental setup

    Systematic evaluations of hyperparameter choices for differential flow motion descriptors establish the following optimal settings:

    • Cell and Block Dimensions: A spatial cell size of 8×88 \times 8 pixels is optimal (reducing to 6×66 \times 6 slightly degrades performance). Fine orientation binning (9 bins) with 2×22 \times 2 blocks (4×4\times overlap) works best for MBH and IMHdiff. For IMHcd, IMHmd, and IMHwd, 6 orientation bins over 3×33 \times 3 cell blocks provide equal performance to 9 bins.
    • Block Normalization: Lowe's hysteresis-thresholded L2L_2 normalization (L2L_2-Hys: L2L_2 norm followed by clipping bin values at 0.2 and renormalizing) significantly outperforms standard L2L_2 and L1L_1-sqrt normalizations for flow descriptors.
    • Derivative Mask Width: Smallest possible finite difference filters ([−1,0,1][-1, 0, 1] with 1-pixel displacement) outperform wider masks (3-pixel or 5-pixel displacements) for MBH.
    • Sub-Pixel Flow Precision: Accurate orientation voting requires deep sub-pixel flow accuracy. Rounding flow field vectors by as little as 1/101/10 of a pixel causes substantial detection performance loss. Coarse block-matching vectors (such as MPEG-4 8×88 \times 8 block vectors) fail to provide sufficient precision and yield inferior detection performance.

Coverage note — Deliberately omitted general introductory reviews of prior pedestrian detectors and background surveys on static Haar wavelets and SIFT.

References

  1. 1.S. Baker and S. Nayar. Pattern rejection. In Proceedings of the Conference on Computer Vision and Pattern Recognition, San Francisco, California, USA, 1996.
  2. 2.D. Comaniciu. An algorithm for data-driven bandwidth selection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 25(2):281–288, 2003.
  3. 3.N. Dalal and B. Triggs. Histograms of oriented gradients for human detection. In Proceedings of the Conference on Computer Vision and Pattern Recognition, San Diego, California, USA, pages 886–893, 2005.
  4. 4.A. Efros, A. Berg, G. Mori, and J. Malik. Recognizing action at a distance. In Proceedings of the 9th International Conference on Computer Vision, Nice, France, pages II:726–733, 2003.
  5. 5.D. Fleet and A. Jepson. Stability of phase information. IEEE Transactions on Pattern Analysis and Machine Intelligence, 15(12):1253–1268, 1993.
  6. 6.B. Galvin, B. McCane, K. Novins, D. Mason, and S. Mills. Recovering motion fields: An evaluation of eight optical flow algorithms. In Proceedings of the ninth British Machine Vision Conference, Southampton, England, 1998. http://www.cs.otago.ac.nz/research/vision.
  7. 7.D. M. Gavrila. The visual analysis of human movement: A survey. Computer Vision and Image Understanding, 73(1):82–98, 1999.
  8. 8.D. M. Gavrila, J. Giebel, and S. Munder. Vision-based pedestrian detection: the protector+ system. In Proc. of the IEEE Intelligent Vehicles Symposium, Parma, Italy, 2004.
  9. 9.D. M. Gavrila and V. Philomin. Real-time object detection for smart vehicles. In Proceedings of the Conference on Computer Vision and Pattern Recognition, Fort Collins, Colorado, USA, pages 87–93, 1999.
  10. 10.I. Haritaoglu, D. Harwood, and L. Davis. W4: Real-time surveillance of people and their activities. IEEE Transactions on Pattern Analysis and Machine Intelligence, 22(8):809–830, 2000.
  11. 11.K.P. Horn and G. Schunck. Determining optical flow. Artificial Intelligence, 17:185–203, 1981.
  12. 12.T. Joachims. Making large-scale svm learning practical. In B. Schlkopf, C. Burges, and A. Smola, editors, Advances in Kernel Methods - Support Vector Learning. The MIT Press, Cambridge, MA, USA, 1999.
  13. 13.B. Leibe, E. Seemann, and B. Schiele. Pedestrian detection in crowded scenes. In Proceedings of the Conference on Computer Vision and Pattern Recognition, San Diego, California, USA, pages 876–885, June 2005.
  14. 14.D. G. Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision, 60(2):91–110, 2004.
  15. 15.K. Mikolajczyk, C. Schmid, and A. Zisserman. Human detection based on a probabilistic assembly of robust part detectors. In Proceedings of the 8th European Conference on Computer Vision, Prague, Czech Republic, volume I, pages 69–81, 2004.
  16. 16.A. Mohan, C. Papageorgiou, and T. Poggio. Example-based object detection in images by components. IEEE Transactions on Pattern Analysis and Machine Intelligence, 23(4): 349–361, April 2001.
  17. 17.C. Papageorgiou and T. Poggio. A trainable system for object detection. International Journal of Computer Vision, 38(1):15–33, 2000.
  18. 18.M. Proesmans, L. Van Gool, E. Pauwels, and A. Oosterlinck. Determination of optical flow and its discontinuities using non-linear diffusion. In Proceedings of the 3rd European Conference on Computer Vision, Stockholm, Sweden, volume 2, pages 295–304, 1994.
  19. 19.R. E. Schapire. The boosting approach to machine learning, an overview. In MSRI Workshop on Nonlinear Estimation and Classification, 2002.
  20. 20.Bernhard Sch¨olkopf and Alex Smola. Learning with Kernels. The MIT Press, Cambridge, MA, USA, 2002.
  21. 21.J. Sun, J.M. Rehg, and A. Bobick. Automatic cascade training with perturbation bias. In Proceedings of the Conference on Computer Vision and Pattern Recognition, Washington, DC, USA, pages II:276–283, 2004.
  22. 22.P. Viola and M. Jones. Rapid object detection using a boosted cascade of simple features. In Proceedings of the Conference on Computer Vision and Pattern Recognition, Kauai, Hawaii, USA, volume I, pages 511–518, 2001.
  23. 23.P. Viola, M. J. Jones, and D. Snow. Detecting pedestrians using patterns of motion and appearance. In Proceedings of the 9th International Conference on Computer Vision, Nice, France, volume 1, pages 734–741, 2003.

Citation

MLA
Dalal, N., et al. “Human Detection Using Oriented Histograms of Flow and Appearance”. Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2006, pp. 428–41, https://doi.org/10.1007/11744047_33.
APA
Dalal, N., Triggs, B., & Schmid, C. (2006). Human Detection Using Oriented Histograms of Flow and Appearance. In Lecture Notes in Computer Science (pp. 428–441). Springer Berlin Heidelberg. https://doi.org/10.1007/11744047_33
Chicago
Dalal, N., B. Triggs, and C. Schmid. 2006. “Human Detection Using Oriented Histograms of Flow and Appearance”. In Lecture Notes in Computer Science. Springer Berlin Heidelberg. https://doi.org/10.1007/11744047_33.
Harvard
Dalal, N., Triggs, B. and Schmid, C. (2006) “Human Detection Using Oriented Histograms of Flow and Appearance”, Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp. 428–441. Available at: https://doi.org/10.1007/11744047_33.
Vancouver
1. Dalal N, Triggs B, Schmid C (2006) Human Detection Using Oriented Histograms of Flow and Appearance. In: Lecture Notes in Computer Science. Springer Berlin Heidelberg, pp 428–441

BibTeX

@inbook{Dalal_2006, title={Human Detection Using Oriented Histograms of Flow and Appearance}, ISBN={9783540338352}, ISSN={1611-3349}, url={http://dx.doi.org/10.1007/11744047_33}, DOI={10.1007/11744047_33}, booktitle={Computer Vision – ECCV 2006}, publisher={Springer Berlin Heidelberg}, author={Dalal, Navneet and Triggs, Bill and Schmid, Cordelia}, year={2006}, pages={428–441} }
Metadata:Crossref

Access the Paper

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

Open PDF