Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
This master thesis relates to machine vision system and object classification. The aim of this paper is to classify the flying objects in images in a computer vision system, for example, an eagle, kite or airplane. In this thesis, large amounts of data will be analyzed and a behavior model will be built for each object as important steps towards improving and automating the object classification system. The application of this thesis is to reduce the deaths of golden and bald eagles due to wind blades.
In this thesis work, a new effective method is presented, namely, a stereo vision system, which is applied in feature selection based on this object classification. Several features are primarily extracted, including the flying height, speed, size and degree of changes in the object parameters.
For image processing and feature extraction, the video acquisition is the first and essential step. Due to the limitation both of equipment and location, the captured videos still do not allow for the collection of sufficient data. For the classification of two objects, a Support Vector Machine (SVM) and Library for Support Vector Machine (LIBSVM) have been employed and implemented in MATLAB. In addition, a preliminary study in relation to the idea of multi-class classification has been conceived and tested by means of an experiment.
In relation to building a behavior model, the various feature properties and characteristics were beneficial with regards to developing the accuracy and robustness of the final classification and recognition results. The results gathered from these two methods in terms of SVM and LIBSVM are compared and analyzed in order to identify their differences and to determine a better solution. Additionally, the possible future work for this project will be discussed.
Results show that 98% of the flying objects can be currently classified by using OVO SVMs and the OVR SVMs. Based on the results of the classification, 85.82% of the flying objects could be predicted correctly.
Key words: machine vision system, object classification, behavior model, stereo vision system, image processing, feature extraction, SVM, LIBSVM, MATLAB