Navneet Dalal and Bill Triggs French National Institute for Research in Computer Science and Control (INRIA) CVPR 05.

Histograms of Oriented Gradients for Human Detection

Navneet Dalal and Bill TriggsFrench National Institute for Research in Computer Science and Control (INRIA)

CVPR 05

OpenCV implement peopledetect.cpp

Introduction Challenge: variable appearance and the wide range of poses Histogram of Oriented Gradients (HOG) are feature

descriptors used in computer vision and image processing for the purpose of object detection.

Basic idea : local object appearance and shape can be characterized rather well by the distribution of local intensity gradients or edge directions.

Similar with edge orientation histograms [4,5], SIFT descriptors [12] and shape contexts [1]

Dataset(1/2)

64x128

Dataset(2/2) INRIA negative images (64x128 samples)

An overview of our feature extraction and object detection chain

http://quyanyun.com/Files/Viso/%E7%AC%AC%E5%9B%9B%E8%AE%B2Dalal-phd-slides.pdf

Person / non-person classification

Implementation(1/7) Color / gamma normalization

o Grayscale, RGB and LAB color spaces optionally with power law (gamma) equalization

o Not obvious effect Gradient Computation

o 1-D point derivatives : uncentred [-1, 1], centred [-1, 0, 1] and cubic-corrected [1,-8, 0, 8,-1]o 3*3 Sobel maskso 2*2 diagonal oneso Gaussian smoothing with σo 1-D at σ =0 work besto The simplest scheme turns out to be the best

DET(Detection Error Tradeoff)

Implementation(2/7) Creating the orientation histograms

o Weighted vote for an edge orientation histogram over cells.o Unsigned gradients used in conjunction with 9 histogram channels

performed best in their human detection experimentso Weight: gradient magnitude itself, or some function of the magnitude (square, square root, clipped)o Gradient magnitude itself generally produces the best results.

cell

Implementation(3/7) Normalization and descriptor blocks

o Owing to local variations of illumination and foreground-background contrast

o Group cells into larger, spatially connected blocks and normalize each block separately

o Two main block geometries : rectangular R-HOG blocks and circular C-HOG blocks.o R-HOG : 3 parameter

• # of cells per block • # of pixels per cell• # of channels per cell histogram• Optimal : 3x3 cell blocks of 6x6 pixel cells with 9 channels.• Gaussian spatial weight

Implementation(4/7) Normalization and descriptor blocks

o C-HOG : 4 parameter• # of angular bins• # of radial bins• The radius of the center bin• The expansion factor for the radius of additional radial bins• Optimal: 4,2,4,2, Gaussian spatial weight is not need

o Block Normalization schemes• L2-norm :• L2-Hys : L2-norm ,clip (limit v<=0.2) and renormalize• L1-norm : • L1-sqrt :

Implementation(5/7)

Implementation(6/7)

R/C-HOG give near perfect separation on MIT database Have 1-2 order lower false positives than other descriptors

Implementation(7/7) Feed the descriptors into some recognition system :SVM classifier

Summary8*8 cell size

Histograms of edge orientationsedge

[-1, 0, 1] gradient filter with no smoothing

8*16 cells 9 unsighted bins=>9 dimension vector

Gaussian spatial window with = 8

R-HOG, 2*2 block size=> 36 dimension vector

L2-Hys7*15 blocks =>descriptor: 3780 dimension vector overlap=1/2

Conclusion We show experimentally that dense grids of Histograms of

Oriented Gradient (HOG) descriptors significantly outperform existing feature sets for human detection.

We study the influence of each stage of the computation on performance.