基於影像辨識路口校正地理座標與行車路徑

Abstract

車輛定位普遍應用在各式各樣的領域中，例如：行車導航、共享機車服務、街景圖資的收集或是自動駕駛，這些應用中，精確車輛定位都是不可或缺的技術。在目前商用解決方案中，全球導航衛星系統（global navigation satellite system, GNSS）是最為廣泛使用的技術，然而，衛星訊號容易受到環境的影響，可能會因衛星訊號被遮蔽或是多重路徑干擾近而影響定位的準確度，我們實驗發現，若僅採用全球定位系統（global positioning system, GPS），在都市地區誤差可能高達50公尺左右。為了彌補定位不準的問題，近年來基於融合各種感測器（如相機、雷達、光達）的技術被廣泛研究。這些方法透過偵測自身週遭的環境及特徵物，並結合已知的地圖資訊以修正GPS誤差。 本文提出一個基於視覺的定位方法，此方法使用單眼相機（monocular camera）及深度神經網路（deep neural network, DNN），其中DNN藉由基於AI的場景理解（context understanding），而非透過檢測和組合各種特徵物來辨識路口（如道路標記及商家招牌這類的特徵，需要透過移動測繪系統（mobile mapping system, MMS）來搜集其地理座標，而MMS成本高且耗時）。藉由降低對單一特徵的依賴，來減輕檢測錯誤造成的潛在問題（例如因為沒有識別到紅綠燈導致最後預測結果為沒有出現路口），此方法也可有效地避免因行車視角不同、道路標記模糊不清或是商家招牌被遮擋造成辨識不到等問題。透過本文所提出的演算法可以低成本且自動化的方法找到地圖上每個路口的地理座標，實驗結果顯示，在各種駕駛環境下，與原始行車接收到的GPS座標相比，我們方法的平均誤差降低了65％，與使用Snap-to-Road相比則降低了48％，所提出的解決方案可以實現小於5米的誤差。

Vehicle localization is commonly used in a variety of fields, such as automotive navigation, scooter-sharing service, collection of street view images, or autonomous driving. To realize these applications, accurate vehicle positioning is an indispensable technology. Among the current commercial solutions, global navigation satellite system (GNSS) is most widely used. However, satellite signals are easily affected by the environment, and the accuracy may be affected due to the obscuration of satellite signals or multipath interference. Our experiments indicated that if only global positioning system (GPS) is used, the error may be as high as 50 meters in urban areas. To improve the inaccurate GPS measurements, sensor fusion-based methods (such as cameras, radars, and LIDAR) have been extensively studied in recent years. These solutions reduce GPS errors by detecting the surrounding features and comparing them with the known map information. In this thesis, we propose a visual-based positioning approach based on a monocular camera and deep neural networks (DNN). When searching and combining various features, such as store signboards and road markers to identify an intersection, geo-coordinates of these features need to be collected through the mobile mapping system (MMS), which is costly and time-consuming. Our DNN uses a context understanding approach to detect intersections. Our approach does not rely on individual features, thus mitigating the potential propagation of a feature-level detection error. For example, instead of using the existence of traffic lights to detect the intersection, we view the entire intersection as an object and detect it directly. In addition, our method can effectively avoid miss detection problems caused by different driving conditions, unclear road markers, or obstruction of store signboards. The method in this paper is evaluated in various driving scenarios. Experimental results show that the average geo-coordinate errors of our proposed method is reduced by 65%, compared to the usage of only GPS, and reduced by 48%, compared to Snap-to-Road. Moreover, the geo-coordinate errors of the proposed low-cost solution can be reduced to less than five meters.