A new classification approach based on geometrical model for human detection in images

Al-N'Awashi, MM 2019, A new classification approach based on geometrical model for human detection in images , PhD thesis, University of Salford.

[img]
Preview
PDF
Download (4MB) | Preview

Abstract

In recent years, object detection and classification has gained more attention, thus, there are several human object detection algorithms being used to locate and recognize human objects in images. The research of image processing and analysing based on human shape is a hot topic due to its wide applicability in real applications. In this research, we present a new shape-based classification approach to categorise the detected object as human or non-human in images. The classification in this approach is based on applying a geometrical model which contains a set of parameters related to the object’s upper portion. Based on the result of these geometric parameters, our approach can simply classify the detected object as human or non-human. In general, the classification process of this new approach is based on generating a geometrical model by observing unique geometrical relations between the upper portion shape points (neck, head, shoulders) of humans, this observation is based on analysis of the change in the histogram of the x values coordinates for human upper portion shape. To present the changing of X coordinate values we have used histograms with mathematical smoothing functions to avoid small angles, as the result we observed four parameters for human objects to be used in building the classifier, by applying the four parameters of the geometrical model and based on the four parameters results, our classification approach can classify the human object from another object. The proposed approach has been tested and compared with some of the machine learning approaches such as Artificial Neural Networks (ANN), Support Vector Machine (SVM) Model, and a famous type of decision tree called Random Forest, by using 358 different images for several objects obtained from INRIA dataset (set of human and non-human as an object in digital images). From the comparison and testing result between the proposed approach and the machine learning approaches in term of accuracy performance, we indicate that the proposed approach achieved the highest accuracy rate (93.85%), with the lowest miss detection rate (11.245%) and false discovery rate (9.34%). The result achieved from the testing and comparison shows the efficiency of this presented approach.

Item Type: Thesis (PhD)
Schools: Schools > School of Computing, Science and Engineering
Depositing User: Malek Maged Al-N'Awashi
Date Deposited: 18 Dec 2019 11:23
Last Modified: 18 Jan 2020 02:30
URI: http://usir.salford.ac.uk/id/eprint/52865

Actions (login required)

Edit record (repository staff only) Edit record (repository staff only)

Downloads

Downloads per month over past year