基於色彩分離與邊界強化的車道線偵測

Abstract

交通安全是生活中關於人身安全的重要議題，隨著近年來交通議題在社會上愈來愈受重視，這也帶動了自駕車及車輛輔助駕駛系統的發展。基於影像處理的車道線偵測系統在自駕車中愈來愈受重視。為了解決車道線檢測系統在行駛中，容易受到附近行進中的車輛影響而檢測錯誤的情況，本論文提出了一種基於影像處理的多車道線偵測方法。首先使用高效率的YOLOP模型過濾掉影像中的車子，接著使用全域的二值化方法找出白色和黃色的車道線，再使用快速的邊緣偵測方法擷取車道線特徵。經由逆透視轉換後，我們使用k-Means來將擷取到的邊緣特徵分類成個別的車道線，最後根據各個車道線類別用拋物線函數來擬合車道線。我們的方法在TuSimple資料集中平均Accuracy、Precision和Recall能達到0.727、0.509和0.407。

Traffic safety is an important topic regarding personal safety in life. With the increasing attention to traffic issues in society in recent years, there is a need for the development of self-driving cars and vehicle driver assistance systems. Image-based lane line detection systems are essential for self-driving cars. In order to prevent the lane line detection from being easily affected by nearby moving vehicles and obstacles during driving, this thesis proposes a multi-lane line detection method based on a series of image processing steps. First, an efficient YOLOP model is used to filter out the cars in the image, and a global thresholding method is used to locate the white and yellow lane lines. Then a fast edge detection method is used to extract the lane line features. After applying the inverse perspective mapping, we employ k-Means to classify the extracted edge features into individual lane lines. Finally, according to each lane line category, a parabolic function is used to fit the lane line. The average Accuracy on the TuSimple dataset reached 0.727, with the Precision of 0.509 and the Recall of 0.407.