Fast Human Detection Using a Novel Boosted Cascading Structure With Meta Stages

Fast Human Detection Using a Novel Boosted Cascading Structure With Meta Stages

Yu-Ting Chen and Chu-Song Chen, Member, IEEE

INTRODUCTIONINTRODUCTION TTECHNIQUES for detecting humans in images have a ECHNIQUES for detecting humans in images have a wide variety of applications, such as wide variety of applications, such as video surveillancevideo surveillance ，， smart roomssmart rooms ，， content-based image/video retrievalcontent-based image/video retrieval ，， andand intelligent transportation systems (ITS) intelligent transportation systems (ITS) 。。

ABSTRACTABSTRACT We propose a method that can detect humans in a sWe propose a method that can detect humans in a single image based on a novel cascaded structureingle image based on a novel cascaded structure.. USE : USE : intensity-based rectangle featuresintensity-based rectangle features ， ， gradient-based 1-D featuresgradient-based 1-D features ，， real AdaBoost algorithmreal AdaBoost algorithm ，， a novel cascaded structure a novel cascaded structure (standard boosted cascade)(standard boosted cascade) ，， meta-stagesmeta-stages 。。

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL Intensity-Based Features:Intensity-Based Features:

:denotes the rth type rectangle feature (r=1~10)

In each block,a feature value can be calculated

and are the Illumination summations in the white and black regions,

:is the feature value in block

The integral-image method is used for fast evaluation of these features,but the results for human detection are notsatisfactory.so add discrimination

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

Gradient-Based Features:Gradient-Based Features: HOG: First, the representation is too complex to evaluate, resulting in a slow detection spe

ed. Second,all the dimensions of a HOG feature vector are employed simultaneously, s

o it is not possible to just use some of them to achieve efficient detection. Third, its computation cost is high since it uses a Gaussian-kernel SVM instead of li

near SVM.

EOH: A EOH feature can only characterize one orientation at a time, and it is represented

by a real value. Many EOH features (with respect to different orientations)can be extracted from an i

mage region, but each feature is only 1-D.

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

EOF Features:EOF Features:

First: The gradient image is calculated from the original image by convolving the edge operator.

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

Second:To compute the EOH features, the pixel gradient magnitude m and gradient orientation θ of each pixel p at location (x,y) in block Bi.

Gx: gradients in the horizontal directionsGy: gradients in the vertical directions

The gradient orientation is evenly divided into K bins over 0 to 180 . The sign of the orientation is ignored; thus,the orientations between 180 to 360 are deemed the same as those between 0 and 180 .

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

Third:The gradient orientation histograms Ｅ i,k in each orientation bin K of block Bi are obtained by summing all the gradient magnitudes whose orientations belong to bin K in Bi.

Fourth: is the feature value of the Kth ( K = 1~ K) EOH feature in block Bi.ε is a small positive value that avoids the denominator being zero.

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

Gradient-Based ED (edge-density) feature :For a block , an ED (edge-density) feature is defined as the average gradient magnitude

is the ED feature value in Bi and ai is the area of BiSimilar to the rectangle features, the integral-image method can be employed for fast evaluation of the ED features.

Combined Feature Pool:r = 1~10, k = 1~K

REAL ADABOOST AND FEATURE POOLREAL ADABOOST AND FEATURE POOL

real AdaBoost algorithm :real AdaBoost algorithm :Given input data z and its feature Value f(z), the weak learner output h(z)

After selecting T weak classifiers,the strong classifier of Real AdaBoost can be expressed as

α is a threshold

A high confidence value implies that the input data is likelyto be a positive sample.

CASCADING FEED-FORWARD CLASSIFIERSCASCADING FEED-FORWARD CLASSIFIERS

A.Contains S stages and Ai is referred to as an AdaBoost classifier in the ith stage.B.In this cascaded structure, detection windows that do not contain humansC.To find an object of unknown position and size in an image usually involves a brute-force search of all possible sites and scales in the image. Since there are usually far more negative windows than positive windows to detect in an image, saving on the detection time of the negative windows increases the overall efficiency of the object detector. DSince more difficult negative examples are used for training in later stages.In the current stage will not be used in later stages.

CASCADING FEED-FORWARD CLASSIFIERSCASCADING FEED-FORWARD CLASSIFIERSTo train a cascaded structure, the goals of the minimum detectionrate of positive examples, di, and the maximum false-acceptancerate of negative examples,fi , are set for each stage Ai.

CASCADING FEED-FORWARD CLASSIFIERSCASCADING FEED-FORWARD CLASSIFIERS

Adding Meta-StagesAdding Meta-Stages : :

“A ” and “M ” denote the AdaBoost stages and meta-stagesMeta-stages: 2-D spaceMi: 2-D vector

CASCADING FEED-FORWARD CLASSIFIERSCASCADING FEED-FORWARD CLASSIFIERS

Meta-Stage Classifier:Meta-Stage Classifier: we choose the linearSVM(LSVM) as the meta-stage cwe choose the linearSVM(LSVM) as the meta-stage classifier because of its high generalization abillassifier because of its high generalization ability and efficiency in evaluation.ity and efficiency in evaluation. ω is the 2-D normal vector of the plane and

β is the offset from the origin. The confidence value of the meta-stage for data is defined as

RESULTSRESULTS

rectangle features = rectangle features = Rec-Cascade (625)Rec-Cascade (625) EOH features = EOH features = EOH-CascadeEOH-Cascade (584) (584) ED features = ED features = ED-CascadeED-Cascade (2492) (2492) a combination of rectangle and EOH features = a combination of rectangle and EOH features = RecEOH-CascadeRecEOH-Cascade

(325)(325) a combination of them as feature = a combination of them as feature = RecEOHED-Cascade (310)RecEOHED-Cascade (310)

RESULTSRESULTS

In our experiments, the maximum FPPW values are about 10^(-3) for most of the cascaded approaches compared. Since there are far more negative windows than the positive windows in an image, a detector shall have a very low false positive rate (e.g., under 10^(-3)), or it might not be practically useful.

miss rate versus false positives per window (FPPW)

HOG add META than RecEOHED-CascadeMetaCascade-2D greatest

RESULTSRESULTS

In our experiments, we consider thefollowing three forms: ,2-D meta classifiers ( n1 = 2, ni = 1 )3-D meta classifiers ( n1 = 3, ni = 2 )4-D meta classifiers ( n1 = 4, ni = 3 )

for a 320X240 testing image, the average processing speeds MetaCascade-2D = 8.61 fps (243)MetaCascade-3D = 8.52 fps (230)MetaCascade-4D = 8.44 fps (201)

HOG-MetaCascade-2D = 6.12 fps (516)

RESULTSRESULTS

HOG-LSVM (0.91 fps)

MetaCascade-2D = 8.61 fpsMetaCascade-3D = 8.52 fpsMetaCascade-4D = 8.44 fps

HOG-MetaCascade-2D = 6.12 fps

RESULTSRESULTS

Thank you for your Thank you for your listening !listening ! 2008.10.212008.10.21