利用基因演算法之Fuzzy ID3方法於階層式場景分析系統

Abstract

本篇論文應用機器學習演算法幫助我們分析影像之場景，並將場景中的物件區域辨識出來。在本篇論文中，我們提出一個階層式場景分析系統，它主要是利用以基因演算法為基礎的Fuzzy ID3理論方法。首先，我們提出一個基於基因演算法的Fuzzy ID3理論方法來產生模糊決策樹，並且從這顆建構的模糊決策樹，萃取出描述資料集的模糊規則。然後我們顯示出這些規則如何應用在車前的場景分析。並且利用自然界的法則，來加以改善場景中物件區塊的修正，以提高辨識的正確性。

本篇論文所提出的階層式場景分析方法，主要是應用在分析從駕駛者的角度往外看出去的道路場景。從測試的結果可以證明，我們所提出的這套系統，可以有效的分出場景中的物件，並且能夠提供於避免潛在車禍碰撞事件的應用。

In this thesis, we consider to utilize machine learning algorithm to segment natural objects in outdoor scene images. First, we proposed a scene analysis system that is rooted from genetic algorithm based fuzzy ID3 method. We develop a genetic algorithm based fuzzy ID3 method, which is designed to generate a fuzzy decision tree and the decision tree can extract fuzzy rules to summarize the regularities existing in the data set. Afterward we show how the resultant rules can be used for object recognition and then apply image ground-truthing to further improve the rule-based object classification accuracy.

The proposed hierarchical scene analysis method is applied to analyze the forward-looking road scene from a car. The testing results have demonstrated the natural object segmentation accuracy is quite high and this method provides a potential application in automated car collision avoidance.