Histograms of Oriented Gradients for HumanDetection

Histograms of Oriented Gradients for Human DetectionNavneet Dalal and Bill Triggs INRIA Rhone-Alpes Grenoble, France

Funding: aceMedia, LAVA, Pascal Network

Histograms of Oriented Gradients for Human Detection– p. 1/1

IntroductionDetect& localize upright people in static images

ChallengesWide variety of articulated poses Variable appearance/clothing Complex backgrounds Unconstrained illumination Occlusions, different scales

ApplicationsPedestrian detection for smart cars Film& media analysis Visual surveillance

Histograms of Oriented Gradients for Human Detection– p. 2/1

Approach& Data SetWe focus on building robust feature sets Classi er is just linear SVM on normalized image windows, is reliable& fast Moving window based detector with non-maximum suppression over scale-space Data set available http://pascal.inrialpes.fr/data/human/ Data SetTrain 614 positive images 1218 negative images 1208 positive windows Test 288 positive images 453 negative images 566 positive windows

Overall 1774 human annotations+ re ections

Histograms of Oriented Gradients for Human Detection– p. 3/1

Feature SetsHaar Wavelets+ SVM: Papageorgiou& Poggio (2000), Mohan et al (2001), DePoortere et al (2002) Rectangular differential features+ adaBoost: Viola& Jones (2001) Parts based binary orientation position histograms+ adaBoost: Mikolajczyk et al (2004) Edge templates+ nearest neighbor: Gavrila& Philomen (1999) Dynamic programming: Felzenszwalb& Huttenlocher (2000), Ioffe& Forsyth (1999) Orientation histograms: c.f. Freeman et al (1996), Lowe (1999) Other descriptors: - Shape contexts: Belongie et al (2002) - PCA-SIFT: Ke and Sukthankar (2004)

Histograms of Oriented Gradients for Human Detection– p. 4/1

Processing ChainOrientation Voting Overlapping Blocks Input Image Gradient Image Local Normalization

Input image

Normalize gamma& colour

Compute gradients

Weighted vote into spatial& orientation cells

Contrast normalize over overlapping spatial blocks

Collect HOG’s over detection window

Linear SVM

Person/ non person classification

Histograms of Oriented Gradients for Human Detection– p. 5/1

HOG DescriptorsParametersGradient scale Orientation bins Percentage of block overlap R-HOG/SIFTCell Block Block

SchemesRGB or Lab, color/gray-space Block normalization,L2-norm, v→ v/ or L1-norm, v→ v/( v1 2 2

v

+

2

+ )

C-HOGCenter Bin

Radial Bins, Angular Bins

Histograms of Oriented Gradients for Human Detection– p. 6/1

PerformanceMIT pedestrian databaseDET different descriptors on MIT database 0.2 Lin. R HOG Lin. C HOG Lin. EC HOG Wavelet PCA SIFT Lin. G ShaceC Lin. E ShaceC MIT best (part) MIT baseline 0.5

INRIA person databaseDET different descriptors on INRIA database

0.2

miss rate

0.1

miss rate

0.1 Ker. R HOG Lin. R2 HOG Lin. R HOG Lin. C HOG Lin. EC HOG Wavelet PCA SIFT Lin. G ShapeC Lin. E ShapeC 10 5

0.05

0.05 0.02 0.02 0.01 10

6

10

5

10 10 10 false positives per window (FPPW)

4

3

2

10

1

0.01 6 10

10 10 10 false positives per window (FPPW)

4

3

2

10

1

- R/C-HOG give near perfect seperation on MIT database - Have 1-2 orders of magnitude lower false positives than other descriptors

Histograms of Oriented Gradients for Human Detection– p. 7/1

Gradient Smoothening& Orientation BinsGradient scale,σDET effect of gradient scaleσ 0.5 0.5

Orientation bins,βDET effect of number of orientation binsβ

0.2 miss rate 0.1 0.05σ=0σ=0.5σ=1σ=2σ=3σ=0, c cor 10 10 10 10 false positives per window (FPPW) 5 4 3 2

0.2 miss rate 0.1 0.05 bin= 9 (0 180) bin= 6 (0 180) bin= 4 (0 180) bin= 3 (0 180) bin=18 (0 360) bin=12 (0 360) bin= 8 (0 360) bin= 6 (0 360) 10 10 10 10 false positives per window (FPPW) 5 4 3 2

0.02 0.01 6 10

0.02 0.01 6 10

10

1

10

1

Using simple smoothed gradients& many orientations helps! Reducing gradient scale from 3 to 0 decreases false positives by 10 times Increasing orientation bins from 4 to 9 decreases false positives by 10 times

Histograms of Oriented Gradients for Human Detection– p. 8/1

Normalization Method& Block OverlapNormalization methodDET effect of normalization methods 0.2

Block overlapDET effect of overlap (cell size=8, num cell= 2x2, wt=0) 0.5

0.1 miss ratemiss rate

0.2 0.1 0.05

0.05

L2 Hys L2 norm L1 Sqrt L1 norm No norm Window norm 10 10 false positives per window (FPPW) 4 3

0.02 0.01 6 10

0.02 5 10

overlap= 3/4, stride= 4 overlap= 1/2, stride= 8 overlap= 0, stride=16 10 10 10 10 false positives per window (FPPW) 5 4 3 2

10

2

10

1

Strong local normalization is essential

Overlapping blocks improves performance, but descriptor size increases

Histograms of Oriented Gradients for Human Detection– p. 9/1

Effect of Block& Cell Size

20

Miss Rate (%)

15 10 5 0 12x12 10x10 8x8 3x3 2x2 1x1

Cell size (pixels) 4x4

6x6

4x4

Block size (Cells)

Trade off between need for local spatial invariance and need for ner spatial resolution

Histograms of Oriented Gradients for Human Detection– p. 10/1

Descriptor Cues

input image

weighted pos wts

weighted neg wts

The most important cues are head, shoulder, leg silhouettes Vertical gradients inside the person count as negative Overlapping blocks those just outside the contour are the most important

avg. grad

outside in block

Histograms of Oriented Gradients for Human Detection– p. 11/1

ConclusionsFine grained features improve performanceNo gradient smooth …… 此处隐藏：3761字，全部文档内容请下载后查看。喜欢就下载吧 ……