透過您的圖書館登入
IP:3.140.185.147
  • 學位論文

利用單眼影像進行可駕駛道路區域分析與車輛偵測

Monocular Vision-Based Drivable Region Labeling and Range Estimation of On-Road Vehicles

指導教授 : 連豊力

摘要


由於感知可以針對有目的性的移動給予資訊,而控制決策則是在可以兼顧及時、準確以及穩健的條件下進行移動,因此在自動學當中,包含機器人學以及輔助系統,感知與控制決策是兩項最重要的課題。 在輔助系統當中,近年來為了避免因為駕駛者沒有專心或是錯誤的判斷所導致的車禍意外,因而提出了利用機器的人工智慧輔助駕駛者的認知判斷以及協助控制的進階安全駕駛計畫。其中,道路偵測是相當熱門的主題,而且已經被廣泛地利用到車道偏離警示系統、車輛偵測與行人偵測系統中;針對不同的功能性,可以將道路偵測的主題分成可駕駛道路區域偵測以及路上的物體偵測共兩種主題。 在可駕駛道路區域偵測的主題中,由於路面的特色很多元不好用人工分析,通常是利用特定的機器學習方法,像是類神經網路和支援向量機器,針對道路區域的部分學習完畢後,在之後的影像中利用學習好的機器認知將路面區域從整張影像中辨識出來。實際上,由於使用這種方法,辨別結果的好壞取決於給機器訓練學習的道路影像,但因為無法窮舉所有可能的道路區域影像讓機器學習,也不可能讓使用者到每種不同的場景時都要再進行一次學習的步驟,因此,這種機器學習的方法並不實際。 而在路上的物體偵測這個主題中,主要都是關於如何成功偵測和辨識路上物體的種類並且進行避障;因此,與路上物體之間的距離估測以及物體的移動路徑的推測對於碰撞時間的估測是相當重要的。 雖然,對於僅提供顏色資訊而沒有提供深度資訊的單眼相機而言,要做距離的估測是一件相當具有挑戰性的事情,但是,我們提出了一個嶄新的方法達成了這個具有挑戰性的事情,此方法包含以下三個步驟,首先是利用從指定的道路區域取得的顏色特徵限制條件,而非利用特定的機器學習機制,進行區域擴張從而找出完整的道路區域的方法,再來是結合在不可行走區域中,利用陰影偵測與車輛結構點進行路上車輛偵測的方法,以及利用相機模型與反透視方法從影像中得到準確的距離資訊,進而建構自車周圍的環境進行認知與危險警示的駕駛輔助系統的方法。 最後,將上述步驟所得到的偵測結果顯示在對於駕駛者而言簡單易懂的格狀的俯視圖中。

並列摘要


Perception and control policy are the keys for automatics, including robotics and human assistance systems. The perception provides information for purposed movements, and the control policy makes the movements in a real-time accuracy-robustness-balanced way. In the human-assistance systems, advanced safety vehicle (ASV) project which support drivers’ recognition judgment and control by the machine intelligence has been proposed to avoid collisions or accidents which are caused by the lack of recognition and miss judgment by drivers in recent years and as known as driving assistance systems. The road detection is a popular topic and it has been widely used to lane departure warning systems, vehicle detection and pedestrian detection. And the road detection topic can be divided into two parts, which are drivable region detection and on-road objects recognition, by different objectives. In drivable region detection, specific machine learning algorithms like neural networks and supporting vector machine are usually used to classify the region of road surface due to the characteristics of road surface are various. But it is not practical for users to train specific machine learning algorithms every time when they are in different characteristic scenes. In on-road objects recognition, the keys are detecting and classifying on-road objects and then avoiding collision. Thus, the distance estimation and prediction of moving trajectories of on-road objects and the host vehicle are important for collision time estimation. Although the distance estimation from a monocular camera is a challenging issue without depth sensors, where a monocular camera can only provide RGB information in pixels, we propose a method including the region growing using color features restrictions estimated from indicated drivable region instead of specific machine learning algorithms, detecting on-road vehicles from non-drivable region by shadow detection and vehicle structure points, and combining a camera model and inverse perspective mapping (IPM) from a monocular camera image to achieve the vehicle surrounding recognition and warning system with accurate distance information for driver-assistance. Finally, the top-view grid map is chosen to represent all the detection results in an easily understood interface.

參考文獻


[1: Siegwart et al. 2011]
[2: Gonzalez & Woods 2008]
[3: Najjar & Bonnifait 2007]
Maan El Badaoui El Najjar and Philippe Bonnifait, “Road Selection Using Multicriteria Fusion for the Road-Matching Problem,” IEEE Transactions on Intelligent Transportation Systems, Vol. 8, No. 2, pp. 279-291, June 2007.
[4: Alvarez & Lopez 2011]

延伸閱讀