由於車輛具有各種引人入勝的特徵需要增強，因此車輛是一個廣泛研究的對象。車輛的重要方面，如方向的定義、距離的計算和速度的確定，在其整體功能中扮演著重要角色。雖然這些特徵可以直接從單眼視頻中獲得，但它們常常面臨校準和遮擋方面的挑戰。在智能交通系統（ITS）研究領域中，車輛速度估計是一個值得注意的挑戰。儘管傳統和深度學習方法在此領域中展示出潛力，但由於實施硬件設備以通過激光雷達、雷達和磁性傳感器等感測器進行數據收集而產生的高昂費用，它們的進展受到了阻礙。在這篇論文中，我們提出了一個模型，該模型由兩個主要組件組成。第一個組件是一個車輛檢測和跟踪模塊，旨在準確檢測和跟踪特定物體，同時解決校準問題。第二個組件是一個單眼視頻中的同態變換回歸網絡，可以在解決單眼視頻中的遮擋問題的同時，高效而準確地估計車輛的速度。通過對多個數據集的實驗評估，我們證明了我們提出的方法優於現有的技術方法，平均均方誤差（MSE）指標的改善幅度約為51.00%。;The vehicle is an extensively studied object due to its various intriguing features that require enhancement. Important aspects of a vehicle, such as defining its direction, calculating distance, and determining its speed, play a vital role in its overall functionality. Although these features can be directly derived from monocular videos, they often face challenges related to calibration and occlusion. Vehicle speed estimation poses a notable challenge within the realm of intelligent transportation systems (ITS) research. Although conventional and deep learning methods have exhibited potential in this domain, their progress has been hindered by the substantial expenses associated with implementing hardware devices for data gathering via sensors like lidar, radar, and magnetic sensors. In this dissertation, we propose a model that consists of two main components. The first component is a vehicle detection and tracking module, designed to accurately detect and track specific objects while addressing calibration issues. The second component is a homography transformation regression network, which efficiently and accurately estimates vehicle speed while addressing occlusion issues in monocular videos. Through experimental evaluations on multiple datasets, we demonstrate that our proposed method outperforms state-of-the-art approaches, achieving a significant improvement of approximately 51.00% in the mean square error (MSE) metric.
