使用反透視映射模型及徑向掃描之單眼視覺障礙物偵測

Abstract

依照交通部的統計資料來看，交通事故造成人員受傷或死亡的數量不斷地上升，其中車輛與其他障礙物碰撞事故發生最為頻繁，因此世界各地有關提升安全性、預防障礙物碰撞之研發與日俱增，很多研究人員提出利用雙眼立體視覺的技術，然而，雙眼立體視覺較為昂貴且需要複雜的安裝程序，因此影像式駕駛輔助安全系統中在移動車輛中以單支攝影機的障礙物偵測是一個具有實用價值與挑戰性的問題，尤其是已提出的現有方法對於靜態障礙物無法有效地偵測。本論文的目標是希望藉由偵測阻礙行進路線上具有一定高度的物體來提升行車安全，此外，藉由標示目標物離本車最近的位置與距離，可有效的避免在倒車上的碰撞。 在本篇論文中，我們提出一個創新的單眼影像式障礙物偵測技術，我們利用有高度的障礙物在於鳥瞰影像具有發散的特性，提出徑向掃描(Radial Scan)的技術來找出障礙物，並且可有效過濾路面標線，解決以往在靜態障礙物上的問題。為了使我們的系統更可靠地偵測障礙物的位置，我們將每次偵測出的狀態於時間軸上進行統計分析。我們提出的系統將會以鳥瞰影像為基礎估計目標物離本車的距離來做更進一步的安全應用，最後我們測試了許多倒車場景和行車於高速公路、一般道路，藉由實驗結果可證明我們提出的系統在障礙物偵測與距離估測上可達到一定的穩健性與準確性。

According to the statistics of the Ministry of Transaction and Communications, traffic accidents caused fatalities and injuries are increasing. Among all of accidents, collisions between vehicles and other generalized obstacles occur most frequently. Hence, researches related to driving safety and preventing obstacle collisions keep increasing. Many researchers proposed obstacle detection techniques based on stereo camera. However, costs and complexities of installing a stereo camera are much higher than those of a single camera. Detecting obstacles via image processing based on single camera is therefore a valuable and challengeable problem. Typically, most of proposed systems applied motion-based methods and hence cannot detect static obstacles effectively without camera motion. The objective of this thesis is to raise the safety of driving by detecting obstacle with certain height. Furthermore, marking the nearest target with measured distance can effectively avoid collision while reversing vehicles. In this thesis, we proposed a novel obstacle detection technique using monocular image. Obstacles diverge and stretch out, while an image is transformed by an IPM (inverse perspective mapping) model. Based on the mapping characteristic, we propose a radial scan technique to locate obstacles on an IPM transformed image. The proposed system can detect static obstacle without false alarming lane makings. Temporal analysis is applied to increase the robustness of the proposed system. Moreover, the proposed system also provides distance information of obstacles via bird-view image. The proposed system was tested for many conditions including parking and driving on the road or highway. Experimental results demonstrated the effectiveness and accurateness of the proposed obstacle detection and distance measurement technique.