路側雷達偵測器之自動化學習演算法研發

Abstract

交通資訊是各類交通應用的基礎, 交通應用包括旅行時間的估計與預測, 壅塞控制, 交通號誌的控制等等. 傳統上, 收集交通資訊的方式是屬於勞力與成本密集, 為了能夠克服傳統收集交通資訊方法的缺點, 本篇研究針對路側雷達偵測所使用的演算法加以開發與改良, 使得路側雷達偵測器能夠提供更正確與更有效的交通資訊（包含車輛記數, 車種分類）. 本篇研究主要是利用高斯混合模型與EM演算法（及其變形）來形成整個路側雷達偵測器的學習演算法核心, 並以現有產品來進行實際道路上的比較測試. 本文針對目前產品所使用的演算法與車輛特徵進行分析, 進而提出更有效的學習演算法來找到正確的車道位置, 使得每一個車道能進行正確的車輛記數. 此外, 每一個車道上的車種（分大車與小車）判斷, 是利用車輛通過偵測區域所得到的反射訊號強度與變化來進行車種學習, 實驗證實亦有八成以上的分類成果.

Traffic information is essential to eRectively perform numerous traffic operations, including travel time estimation or prediction, congestion control policies, and traffic signal control strategies. Traditionally, gathering traffic information is extremely labor and cost intensive, meaning the assistance of advanced information technology equipment is crucial. For overcoming the defects of traditional detectors, the road-side radar detector is adopted in this study, and more advanced learning algorithms are developed to achieve the automation and enhance the accuracy. The basic traffic information includes traffic count, vehicle classication and vehicle speed estimation. Counting traffic in a single lane is a basic task that can be achieved by using traffic detectors to detect passing vehicles, but it is difficult for road-side radar detectors to simultaneously detect diRerent vehicle types in multi-lane environments, because the signals reflected from passing vehicles in a single lane influence neighboring lanes. The spread of reflected signals created difficulty in accurately identifying lanes. Hence, this study first develops a learning procedure for road-side radar detectors to form an on-line traffic lane estimator. An on-line traffic lane estimator is modelled by Gaussian mixture model (GMM) based on span and conflict information and trained by using the proposed variant of expectation maximization (EM) algorithm. The numerical results demonstrate on-line traffic lane estimator can work well in real-world scenarios, and the accuracy of traffic lane estimator is verified by counting traffic in diRerent lanes. Besides, the real-time vehicle classifier for road-side radar detectors in multi-lane environments is for the first time presented. A two-dimensional Gaussian Mixed Model is employed to develop the learning model based on FMCW radar data. An EM algorithm is thus implemented to maximize the likelihood of the formulated learning model; consequently, the model could be used for classifying small and large vehicles in multi-lane environments simultaneously, so that traffic information can be obtained at a relatively lower cost. In the suburban field test, the accuracy of real-time vehicle classifier in multi-lane environments can achieve more than 88%.