本篇論文中,我們提出了一個能夠適用於不同低解析度道路監控攝影機的車輛分類方法。在智慧型運輸系統中,都市環境下的交通分析一直是這塊研究領域中的重要議題。除了車流分析之外,車輛種類也是智慧型運輸系統中重要的資訊。在本論文中,我們將車輛區分成了四個種類,公車與其他大型車、貨卡車、轎車、摩托車。我們將深度學習方法(VGG16)應用於低解析度之車輛樣本分類,並與傳統的機器學習方法(SVM)進行實驗比較結果。此外,我們透過車輛追蹤來校正單一畫面的分類結果,將同一輛車在時間軸上的所有分類結果進行投票,以決定最後的分類結果。實驗結果顯示出不論在訓練過或未經訓練過的監視器畫面皆能取得約七成以上的正確率。
Vision-based vehicle detection and classification in urban area plays an important role in the development of intelligent transport systems(ITS). In Taiwan, most existing traffic cameras capture very low resolution and poor quality videos for surveillance. It becomes a challenge problem for classifying the types of vehicles by traditional machine learning scheme in such traffic videos. In recent years, we are observing an accelerated success of deep learning in most image recognition tasks. In this study, we classify vehicles into four categories including scooter, car, truck and bus. In order to understand the accuracy of vehicle classification in practice. We apply support vector machine (SVM) with handcrafted features and VGG16 architecture to compare the performance of traditional machine learning with that of deep learning on low resolution traffic videos. We also propose a vehicle tracking algorithm to collect the classification result of the same vehicle in individual frames, to determine the type of vehicle by voting. The experimental results show that the classification accuracy of VGG16 is better than that of SVM. It achieves over 70% for all the four categories of vehicles in our test video set.