SVM as the last cascade. The training and test procedures are shown as follows. MSO and HOG feature extraction Ada boost and SVM class ifiers trai ning Fine c lassificat ion by SVM Pedestrian models
J;= F(arctan(Dx/ Dy»
(2)
Detection result
The rest of this paper is organized as follows. Detection method is presented in section 2. Experiments and conclusions are presented in the section 3 and section 4.
F is a function that can convert the continuous angle values into discrete feature bins in{a, 1,...,8}. Here we use 9 feature bins by classifying the[0,180] scope. MSO feature set: We use a boosting procedure to search and combine basic units from a candidate sets that are formed by shifting units of different size (from 6x6, 8x8 to 20x20 pixels) with 2-pixel interval, obtaining totally 11,180 block units- which are similar to Haar-like features . The difference among them is that the former is limited to square blocks instead of rectangles, since the orientation is calculated on a square image block, not on a rectangle one. The pattern that appears like rectangle (e.g., arms, legs, etc.), which can be reflected by assembling mu
ltiple square blocks . Histogram of po s iti ve s
2. FEATURE EXTRACTION AND PEDESTRAIN DECTECTION In this section, feature extraction, selection, classifiers training, and pedestrian's detection are presented.
H. I
2.1. Feature extraction Basic MSO units and coarse orientation features: The basic unit in a fixed scale is a square block as shown in Fig.2a-b. Calculating the unit coarse feature contains two steps. I) Calculating unit orientation, 2) quantizing the orientation into feature bins . The square block is first divided into left-right sub-blocks as shown in Fig.2a, then the color value is integrated in the sub-blocks to obtain the unit horizontal gradient by the following equation: dx=
(a)
(b)
Fig.3. Unit selection (a) illustration of weak classifier (b) units of a lO-stage adaboost To evaluate the MSO unit, we need firstly to construct a weak classier as C.= I
{a,
if T/<5, J;<5, T/ 1, otherwise
(3)
I XEL<ft~bbIOC/(X) - XERigt;,bbIOC/ (X )1
(1)
C, is the classification result on a unit orientation feature,
leX) is the pixel value of position X . For the three color
r;' and r;2 are two thresholds for the weak classifier, which are determined in the learning process. The reason why we use a scope of orientation as a weak classifier is that all the local contour of a pedestrian is in a scope. For example, the orientation of shoulder is about from 0 to 45 degree, and that of legs is from about 45 to 100 degree. Given the weak classifier in (3), a standard boosting procedure is adopted to select coarse unit set. All selected features construct the MSO feature set. HOG features: The proposed MSO feature, a kind of patch-based image descriptor, is selective for target shape, and can construct object contour model. But such descriptors are limited in their ability to capture variance in the object appearance. HOG feature is histogram based descriptor, which contain the statistical information of the image patch, and have been shown to be very robust with respect to object transformations. HOG features and SVM classifiers are used to train the fine classier following Dalal's work[9]. A 64 X 128 training sample is divided into cells of size 8 X 8 pixels and each group of 2 X 2 cells is
channels (R, G or B channel), the channel of largest dx value is selected. And a similar operation is used to calculate vertical unit gradient dy . The calculation of (I) is speeded up with"integral image"[5]. . 1
20x20 (a) (b) Fig. 2. Feature calculation on MSO unit, (a) Multi-scale unit and the orientation calculation, (b) Multi-scale units 8x8 We can then calculate the coarse orientation of the unit with:
-=-rn EEl C==~ 6X6
EE
551




Detection using a Boosted Cascade of Simple ...cascade classifiers Discard the background regions ...applications for a extremely fast face detector ?...
Detection using a Boosted Cascade of Simple ...cascade classifiers Discard the background regions ...applications for a extremely fast face detector ?...
热门文档