Detection of partially occluded human using separate body parts classifiers

Nurul Fatiha , Johan (2015) Detection of partially occluded human using separate body parts classifiers. Masters thesis, Universiti Teknikal Malaysia Melaka.

[img] Text

Download (1MB)


The application of computer vision in the surveillance system has provided huge advantages in the field of security and safety system. In recent years, human detection and classification subjects have shown an increasing focus in finding specific individual such as in the case of detecting person in crowded places at a time. Detection and classification of human can be a challenging task due to the wide variability of human appearance in terms of clothing, lighting conditions and the occlusion. These constraints directly influence the effectiveness of the overall system. To cope with these problems, human detection and classification system is presented in this thesis which requires fast computations in addition of accurate results. The propose system will first detect the human in an image by using YCbCr color thresholding for skin color detection algorithm and then classify the body parts using artificial intelligent neural network classifier into specific class and finally extend the classification system with the majority voting technique in order to improve the classification performance.The first hypothesis of the research is that YCbCr skin color detection method can be used to detect and identify the exposed human body parts even with the existence of various illumination conditions and complex background. In this work, the body parts then only cover face and hands. The body features are then extracted using feature extraction technique with the dimension of region detected fixed to a standard size.These body features are then used as an input to neural network system in order to classify the body parts into specific class. Meanwhile each class consists of three classifier which is taken from the extracted body regions and separated into face classifier, right hand classifier and left hand classifier. Finally, the results of each body parts classification will be processed using majority voting technique for overall conclusion of the classification system which is robust to partial occlusion. Experimental results indicate that the human detection using YCbCr color space is capable to detect the human body with the percentage of face detection is 92%, right hand detection is 86% and left hand detection is 85%. Meanwhile the performance of ANN classification system is successful in identifying face, right hand and left hand which are 90%, 73% and 74% respectively. Whereas, the accuracy of all 9 classes (Class A until Class I) is found to be 43% and highest to be 95%. Based on the extended classification system using majority voting technique, the results have shown a bit improvement on the classification performance for all 9 classes which is the lowest is increase to 45% and the highest is increase to 100%.

Item Type: Thesis (Masters)
Uncontrolled Keywords: motion perception (vision), computer vision, human mechanics
Subjects: T Technology > TA Engineering (General). Civil engineering (General)
Divisions: Library > Tesis > FKE
Depositing User: Norziyana Hanipah
Date Deposited: 25 Jan 2016 01:24
Last Modified: 25 Jan 2016 01:24
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item