基於影像追蹤技術偵測隧道內變換車道行為

Abstract

本研究主要透過影像處理技術實現智慧型交通監控系統，目的在於偵測變換車道事件於隧道環境內。系統主要分為車輛偵測、車道軌跡模型統計、車輛追蹤、違規變換車道事件偵測。本篇論文利用非傳統的Sobel正向邊緣來追蹤車輛底部，使系統在光線複雜的隧道仍能正常的追蹤車輛，並能有效地降低陰影及車燈等光影變化對背景法的影響，另外追蹤車底也能解決車輛於真實世界3-D座標投影到影像2-D座標所損失的深度資訊造成所屬車道的誤判。 在演算法中，首先我們會先統計車輛在車道上行駛的情形，得到車行軌跡資訊，經過線性迴歸後，得到一個線性變化的車輛大小，以期能使追蹤軌跡更穩定，執行時就以這個資訊做為參考；系統開始執行時使用背景法以及移動的正向邊緣來偵測車輛，在追蹤時輔以卡爾曼濾波器來修正追蹤軌跡，這裡的卡爾曼濾波器是設計使車輛追蹤的結果能夠更平滑，以方便辨識軌跡。最後以貝氏檢測檢查變換車道的軌跡是否正確。以降低大車及錯誤追蹤所產生的誤報。 實驗場景為一三車道隧道口，選擇隧道口的原因在於這個位置是單白虛線，在此可以捕捉到許多變換車道的行為，以驗證演算法。演算法在經過測試之後不僅在車輛計數有很高的準確率，對於變換車道的偵測也有不錯的偵測率。

The paper aims to detect vehicle which change lanes inside the tunnel. The system consists of vehicle detection, lane model trajectory statistics, vehicle tracking, and illegal lane changing event detection. To have more accurate data and statistics, a positive edge method of Sobel is chosen to track the bottom of vehicles. This method can effectively reduce shadows and lights and other lighting effects which may change the Background subtraction algorithm. In addition, tracking bottom of the vehicles can also solve the problem caused while projecting real-world 3D coordinate onto 2D coordinate image. In terms of algorithm, first we collect vehicle travelling data; put it through the linear regression to obtain the linear variation in terms of the size of the vehicle. The linear variation statistics can help stabilize the tracking; it is also used as a reference. The system begins with the Background Law and moving forward edge to detect moving vehicles, while combined with Kalman filter to amend tracking. The Kalman filter is designed to help smooth the tracking result, to facilitate the identification of the trajectory. Lastly, the result was re-tested by the Bayesian testing, to reduce the error tracking, and false alarm. The experimental scene was set at a three-lane tunnel. The reason for this is that the lane at the entry is single white dotted line (where vehicle is permitted to cross to different lane); you can easily capture the behavior and numbers of the lane changes. After tested the algorithm, it not only have a high accuracy detecting numbers of the vehicle, the lane-change rate also have accurate rating.