Obstacle Detection and Direction Decision for Monocular Visual Based Robots

現今機器人與行車安全監控電腦等是相當熱門的議題，很多研究利用多樣感應器研究用於監控周圍環境保護載具安全，其又以直接使用電腦視覺最為熱門。有相當多的雙視覺研究，建立立體環境空間來偵測路徑上的障礙物。但單視覺障礙物偵測通常能有更低的建造成本與更簡易的架設方式，使用者不需高檔設備或精密安裝即有不錯效果。本文提出一種基於單視覺電腦視覺與LK 光流法(Lucas–Kanade Optical Flow Method)，使用一般webcam 並輸入架設參數，建出地面視野的座查表，再即時依照畫面中物體移動模式來判斷使否有不在地面視野帄面移動的物體，並於每次偵測時都能一併依危險特徵點位置決定路徑方向的方法。Nowadays researches on robotics and vehicle safety are very popular.Various researches use multiple sensors detecting surrounding environment to protect vehicles. Lots of researches detect obstacles by using stereo vision to create three-dimension environment, nevertheless monocular vision is of lower cost and easy setup, and user can get great result without advanced equipment and accurately installation. This thesis proposes a method based on monocular vision, Lucas-Kanade optical flow, and a normal webcam with few setup parameters, to create a horizon plane. By Observing objects in sight and check whether they are in the plane or not, we can obtain a safe path for vehicle.1. 緒論...................................... 5 1.1 前言..................................... 5 1.2 研究動機................................. 8 1.3 研究貢獻................................. 9 1.4 論文架構................................. 9 2. 相關研究.................................. 10 2.1 色相分離法障礙物識別..................... 10 2.2 特徵選取................................. 11 2.3 特徵追蹤................................. 13 3. 研究方法.................................. 19 3.1攝影機拍攝畫面空間座標推導................ 19 3.2區別安全特徵點與危險特徵點................ 24 3.3判定畫面安全方向.......................... 29 3.4演算法整合................................ 34 4. 實驗結果.................................. 35 4.1 軟硬體設置............................... 36 4.2 測試結果................................. 37 5. 結論與未來方向............................ 40 6. 參考文獻.................................. 4