Three-dimensional Object Recognition In Video Tracking System

Technique of three-dimensional object recognition is one of the most important techniques in video tracking system and also a difficult problem in it. The thesis begins with the theory of object recognition, introducing its history and situation, and then compares with the superior methods, at last advances a kind of project that combines moment invariants and artificial neural networks. At length, the thesis concludes the parts of image processing, moment invariants calculation and recognition.Firstly, in the stage of image processing, I gave the object images an effective operation. Because of disadvantage of the usual medium filter, I used a new and quick medium filter, and then I compared among four methods of edge detecting and chose the Prewitt way, receiving clear and complete edge without much noise. Moreover I got an automatic program to divide the images into two gray-levels quickly and effectively.Secondly, in the stage of feature abstracting, I discussed the concept of feature and abstracted moment invariants as the feature of images because they can reflect the feature of images and keep the wonderful invariant character to moving, revolving and stretching. At last I normalized them to reduce the calculation.Thirdly, in the stage of recognition, I introduced the concept and character of artificial neural networks, and then chose self-organizing feature map as my recognizing method considering its powerful self-learning and recognizing ability. At last, I carried out the networks fit for my research with the program named Kohonen. It worked well.This project comes true at the platform of VC and MATLAB. From the experimental data through two plane models, the system can recognize the target at the rate of eighty six point five seven percent and has the general recognizing ability.