利用自動道路標記偵測的車道定位

Abstract

車道定位系統在未來智慧城市中扮演重要的角色，透過辨識車輛位於哪個車道，可以延伸出各式各樣的應用，例如：車道維持輔助系統、車道變換輔助系統和科技執法。現有的車道定位系統通常只依靠道路上車道標線來定位，例如:非基於影像的方法像是光學雷達(LiDAR)藉由向環境照射光束以及接收反射的訊號，偵測車道線位置並重建3D的模擬圖，得以定位車道，此種方法雖然準確但是比基於影像的方法昂貴許多；基於影像的方法則是透過分析影像中的車道線來定位，此種方法能有效地處理像高速公路或快速道路等簡單的環境，但是在城市中容易因為路面上的道路標示而產生誤判。 本文提出一個融合基於傳統電腦視覺的車道線偵測方法以及基於深度學習的道路標示偵測模型，來實現車道定位。其中，使用道路標示偵測的模型來濾除因道路標示而誤判的車道線，再透過消失點偵測 (vanishing point estimation)及路寬資訊濾除掉因雜訊而誤判的車道線，包含：道路分隔島、人行道、陰影邊緣等，最後透過移動平均濾波器 (moving average filter)，解決因車道虛線而導致沒有偵測到的問題。 為評估此法在實際運用中的系統效能及可靠度，我們在各種不同的駕駛環境中進行測試，例如:晴天、陰天、雨天等。實驗結果顯示，與基於深度學習的模型相比，本篇所提出的方法在精確度上提升了8％。此外，此法與基於深度學習的模型相比，辨識速度提升了100％。該深度學習的模型會針對圖中的所有像素做預測，跟道路標示偵測的模型相比提高了計算量，我們相信本文所提出的方法相對於使用深度學習的模型更適用於計算資源較為不足的裝置。

The lane-level positioning system plays an important role for smart cities in the future. By knowing which lane the vehicle is in, a variety of applications, such as lane-keeping assistance, lane change assistance and traffic rules compliance can be realized. Existing lane-level positioning systems only rely on lane lines on the road. For example, the non-image-based approach such as light detection and ranging (LiDAR) localizes the vehicle by detecting the positions of lane lines and reconstructing the 3D structure of surroundings through emitting and receiving laser signals. This approach gives high accuracy. However, it is costly compared to image-based approaches. The image-based approach finds the position of the vehicle by analyzing the lane lines in the image. This approach is efficient when dealing with simple environments such as highways and expressways. However, it gives wrong results due to road markings in cities. In this thesis, we proposed a lane-level positioning method which fuses a traditional computer vision-based lane detection method and a deep-based road markings detection model. In the proposed method, we use the road markings detection model to filter out false detected lane lines due to road markings. Moreover, we exploit the vanishing point estimation and lane width to further filter out false detected lane lines from the edges of refuge islands, sidewalks and shadows on the road. Finally, we use a moving average filter to solve the miss detection problem due to dashed lane lines. To evaluate the performance and robustness of the proposed method, we conduct experiments in various driving scenarios, including sunny, cloudy and rainy days. Experiment results show that the proposed method increases the precision by 8% compared to a pure deep-based model. Furthermore, the proposed method is twice as fast as the deep-based model. The deep-based model predicts labels for all pixels in images, which increases computational cost compared to the road marking detection model. We believe that the proposed method is more suitable for devices with limited computing power than the deep-based model. In this thesis, we proposed a lane-level positioning method which fuses a traditional computer vision-based lane detection method and a deep-based road markings detection model. In the proposed method, we use the road markings detection model to filter out false detected lane lines from road markings. Moreover, we exploit the vanishing point estimation and lane width to filter out false detected lane lines from edges of refuge islands, sidewalks and shadows on the road. Finally, we use a moving average filter to solve the miss detection problem due to dashed lane lines. To evaluate the performance and robustness of the proposed method, we conduct experiments in various driving scenarios, including sunny, cloudy and rainy days. Experiment results show that the proposed method increases the precision by 8% compared to a pure deep-based model. Furthermore, the proposed method is twice as fast as the deep-based model. We believe that the proposed method is more suitable for devices with limited computing power than the deep-based model.