利用多台KINECT裝置及自動車作園區安全巡邏之研究

Abstract

本研究提出了一個利用有視覺之自動車應用於園區安全巡邏之系統，該系統利用搭載了多台KINECT攝影機裝置的自動車當作實驗平台，能在KINECT提供之3D影像中計算出實際物體的立體資訊。首先，使用者可利用建立的環境學習介面來記錄自動車航行地圖及相關參數，此地圖包含自動車在戶外環境中行走的路徑、路徑中學習之路標的位置與學習順序，而相關的導航參數(KINECT編號、彩色/深度資訊、偵測區域定義等等…)可輔助自動車完成航行。接著，將學習的路標(燈座、消防栓、下坡道、樹幹)作為定位點以輔助自動車進行導航，並且於定位成功後將自動車移動至學習的位置上以縮小航行中的誤差。我們整合上述兩項技術並提出一個擁有自動定位和導航功能的自動車系統。此外，本研究也提出了一個新的導航概念，可利用彩色或深度影像做物體辨識，再利用3D影像計算物體實際位置並調整自動車位置以達到自動車定位與位置校正的目的。接著，我們提出了一個使用深度影像做線段跟隨的技巧，利用此技巧可讓自動車於人行道上航行並且與道路保持適當的距離。最後，本研究提出了幾個使用深度影像偵測路標並且定位的方法，並且改善了SURF演算法於深度影像中匹配的結果。實驗結果顯示本研究所提方法完整可行。

A vision-based autonomous vehicle system equipped with KINECT devices for security patrolling on sidewalks in outdoor environments is proposed. A small-size vehicle with three onboard KINECT devices is designed to build the system. At first, a learning procedure is proposed for a trainer to guide the vehicle to extract specific features, including navigation path, color/depth information, pre-selected landmark objects, and vehicle location with respect to the landmark. Next, a strategy of vehicle navigation with a line-following capability is proposed, by which the vehicle may be guided to navigate according to the node data of the learned path, detect along-path landmarks using SURFs, and match the features with the learned data based on the measures of contrast difference and Euclidean distance. In addition, a vehicle location estimation technique for path correction utilizing the landmark matching result is proposed, which is based on the use of an ICP algorithm with the depth information as input according to the criterion of minimum MSE. Furthermore, techniques of ramp and curb-line detection have also been proposed, both for use to guide the vehicle safely on the path as well as to provide environment features and adjust the vehicle orientation. Good experimental results show the flexibility and feasibility of the proposed methods for the application of security patrolling in outdoor environments.