使用HOG-SVM架構的交通標誌辨識

Abstract

近年來駕駛人輔助系統逐漸成為現代車輛的流行配件，而且對將來的自主車輛來說也是不可或缺的元件。一個進階的駕駛人輔助系統可以自動偵測環境並且提供導航。自動交通標誌辨識為一個用在進階駕駛人輔助系統的科技。交通標誌有一些特徵使得它們能簡單地被人們偵測與辨識。它們有著特定的形狀與顏色。而且它們通常會垂直地放置在一定的高度並面對著來車的方向，所以在拍攝的影像中它們的旋轉及幾何變化更能被預測。 在這篇論文裡，我們提供了一個交通標誌辨識的方法，使用被很多的物件辨識程序證明為有效且運算上也很有效率的兩種特徵，分別是Histogram of Oriented Gradients (HOG)跟Gabor特徵。我們採用線性SVM做為分類的方法，使用這兩個影像特徵的性能進行模擬與比較。 我們也減少特徵數量以加速運算的時間，發現在不影響準確率太多的情形下，運算時間能顯著地降低。在一些情形下，適當地減少特徵甚至能增加準確率。我們研究了成功與失敗的例子以了解其原因。我們的觀察成為本論文中不可或缺的一部分。收集足夠數量的訓練資料是機器學習系統的一個瓶頸。因此，我們基於一些收集到的樣本去合成訓練資料。我們用真實的測試資料評估這種方法的性能。到目前為止，我們的合成訓練資料產生的系統效果較差。

In recent years, driver assistance system becomes an increasing popular component of a modern vehicle and its elements are essential for future autonomous car. An Advanced Driver Assistance System (ADAS) can detect the environment and provide navigation automatically. One technology used in ADAS is automatic traffic sign recognition. Traffic signs have a few features so that they can be easier to detect and identify by human beings. They have specific shapes and colors. And they are usually placed vertically at certain height and facing the incoming car direction. Thus, their rotation and geometric variations in the captured images are more predictable. In this thesis, we propose a traffic sign recognition method, using two features which have been demonstrated effective in many object recognition applications and are rather efficient in computation. They are the Histogram of Oriented Gradients (HOG) feature and the Gabor filter feature. We adopt a linear support vector machine (SVM) for object classification. The performance of using these two image features are simulated and compared. We also reduce the feature number to improve the computing time. We found that without affecting the accuracy much, the computation time can be reduced significantly. In some cases, the proper feature reduction can even increase the accuracy. We study the successful and failure cases to understand the reasons. Our observations become an indispensable part of this thesis. To collect sufficient amount of training data is one bottleneck of a machine learning system. We thus synthesize the training data based on a few collected samples. We evaluate the performance of this approach on the real test data. So far, our synthesized training data produce less effective system.