Information Technology Reference
In-Depth Information
either based on pixel or on feature [4]. The feature based learning is beneficial than
pixel based learning, because it is difficult to train finite quantity of data using pixel in
comparison to encoding of ad-hoc knowledge using feature.
To address the human object classification, various researchers proposed their so-
lutions. Some approaches are based on pixel based method whereas some approaches
are based on feature based method. Sialet et al. [5] used Haar like features along with
the decision tree in their pedestrian detection system. Viola and Jones [6] used modi-
fied version of Haar basis function for object detection. Dalal and Triggs [7]
proposed Histogram of Oriented Gradient (HoG) as a feature descriptor for object
detection. Cao et al. [8] proposed a method by extending the Histogram of oriented
Gradient, to boost the HoG features. Lu and Zheng [9] proposed a visual feature for
object classification based on binary pattern. These visual features are rotation invari-
ant and exploits the property of pixel pattern. Lowe [10] used Scale Invariant Feature
Transform (SIFT) as a feature descriptor for object recognition. All the methods dis-
cussed above depend on only one feature evaluation set therefore they have some
local advantages depending on the features used.
Yu and Slotine [11] proposed a wavelet based method for classification. Method
proposed by Yu and Slotine [11] uses real valued wavelet transform, but real valued
wavelet transform is not suitable for surveillance application, because in case of video
surveillance, object of interest may be present in translated or rotated form among
different frames and coefficients of real valued wavelet transform corresponding to
object region changes abruptly across different frame [12,13]. Use of complex wave-
let transform can avoid this shortcoming as it is shift invariant in nature.
Motivated by work of Yu and Slotine and properties of complex wavelet transform
Khare et al. [2] proposed human object classification method by using Dual tree com-
plex wavelet transform. Combining two or more features in one is the recent trend and
give more accurate results in comparison with use of single feature based object clas-
sification. Here we are extending our earlier work [2], by proposing a new method for
human object classification based on combination of Dual tree complex wavelet trans-
form and Zernike moment as a feature set. We have used support vector machine
(SVM) classifier for classifying human and non-human object classes. The Dual tree
complex wavelet transform having advantages of shift invariance and better edge
representation as compared to real valued wavelet transform. Zernike moment also
have many important properties such as translation invariance, rotation invariance etc.
We have experimented the proposed method at multiple levels of Dual tree com-
plex wavelet transform. We have also compared the proposed method with the meth-
od using coefficients of real-valued discrete wavelet transform as a feature set. We
have compared the proposed method with other state-of-the-art methods proposed by
Khare et al. [2], Dalal and Triggs [7], Lu and Zheng [9], Renno et al. [14], and Chen
et al. [15] in terms of confusion matrix. We have taken three different performance
metrics: average classification accuracy, true positive rate (recall), and predicted posi-
tive rate (precision).
Rest of paper is organized as follows: Section 2 describes basics of features (Dual
tree complex wavelet transform and Zernike moment). Section 3 describes support
vector machine classifier and section 4 describes the proposed method. Experimental
results, analysis and comparison of the proposed method with other state-of-the-art
methods are given in section 5. Finally conclusions of the work are given in section 6.
Search WWH ::
Custom Search