- 學位論文
倒傳遞神經網路與地平面立體視覺作用於自動車導航之障礙物偵測與道路分類
Obstacle Detection and Road Classification Applied to Navigation of Automatic Land Vehicle Based on Back-propagation Neural Network and Ground Plane Stereo Techniques
摘要
在自動車導航領域中利用影像做環境場景的識別是十分重要的。本篇論文中利用對地平面校準後之左右影像，使校準後影像相減後地面視差為零，再由校準後影像做區塊匹配對應，取得視差圖，圖中我們可以清楚地分離地面與立於地面障礙物，並推算出障礙物深度與高度資訊，另外為使環境資訊更加充足，我們將從原始影像中取得各類型道路的顏色、紋理等特徵，再經由倒傳遞神經網路進行訓練和分類，分類的平均辨識率達85%以上，提供自動車找到可行的路面區域並做行駛的選擇，由於分類結果只是區塊性質相似，藉由3D定位資訊來驗證影像分類是否誤判，並利用高度、深度及多類道路資訊實現自動車導航系統。此研究透過PIC控制自動車平台，並由電腦端下達指令。車子控制系統分兩部分，一個由伺服馬達控制轉向部分，另一個則由直流馬達控制速度。當程式運行時，速度可達每秒15張，透過實驗可讓自動車在校園行走，證明所提方法的可行性。
並列摘要
In this thesis, we present an approach for environment recognition in automatic land vehicle (ALV). By using block matching for the two ground plane alignment images, we can obtain the disparity map. In the map, we can clearly detect obstacles from ground plane, and calculate the height and depth information of the obstacles. To sufficient environment information, we can obtain the features of the road (like colors, textures, etc.) from the original image. By the environment information obtained, and using back-propagation neural network (BPNN) to train and classification, the ALV will be able to recognize drivable surface, with a recognition rate of 85%, providing various options for navigation. Due to classification results are only template similar, using 3D information and road features to verify classified image will be able to achieve ALV navigation system. In this study, ALV platform is controlled by a PIC controller, and the controlling instruction is sent from a computer. The motor control system is divided into two parts: one for controlling the direction derived by a servo motor, and the other for controlling the speed of going forward or backward with a DC motor. The proposed ALV system can reach at a rate of 15 frames per second. The ALV system has been performed in NTUT campus to demonstrate the effectiveness of the proposed method.