- 學位論文
多重特徵車輛防碰撞方法設計
Development of a Vehicle Anti-collision Method Exploiting Multiple Features
摘要
近年智慧型運輸系統(Intelligent Transportation System,ITS)的蓬勃發展，自動車(Autonomous Guided Vehicle, AGV)的開發更為矚目，隨著行車安全議題受到重視，車輛的主被動式安全輔助已成為車用電子領域重要的研究議題，全球每年的交通事故居高不下，交通意外多為疲勞駕駛或車輛操作不當，故本研究提出多重特徵車輛防碰撞方法設計，達到碰撞預防。 Viola & Jones [22]提出具有分層偵測器(Cascade detector)概念的Adaboost演算法，近年廣泛運用於物件偵測，訓練輸入使用車輛影像，當基於Adaboost訓練的正負樣本數量龐大與車輛偵測準確度到達訓練瓶頸時，如何再使偵測的誤報情況減少且提升偵測的預測率(Precision rate)與準確率(Accuracy)為本研究重點。 隨著影像處理技術蓬勃發展和低成本攝影機的推陳出新，車輛配備皆包含各類影像紀錄器，本研究提出一個使用單鏡頭影像處理的多重特徵車輛防碰撞方法設計，此設計核心演算法基於Adaboost結合Haar特徵再加入多重車輛特徵，包含了陰影特徵、邊緣紋理特徵、尾燈特徵、道路特徵，其目的為提升車輛偵測預測率(Precision rate)與準確度(Accuracy)，降低車輛偵測誤報情形，本研究技術包含四個階段: 影像前處理(Pre-Processing)、車道偵測(Lane Detect Processing)、車輛偵測(Vehicle Detection)、車距警示(Vehicle Distance Warning Processing)，影像前處理經由縮小感興趣區域，降低Adaboost偵測的運算量提升效能，車輛偵測第一階段Adaboost偵測車輛候選區，第二階段多重車輛特徵濾除Adaboost錯誤偵測候選區，多重特徵僅在候選區內運算，具有低運算量與低誤判的優勢，最後針對車道偵測定義出主車道，估計出主車道前車車輛距離，智能的多段碰撞預防警示，達到車輛碰撞預防。
並列摘要
Given the rapid expansion of autonomous guided vehicle ownership worldwide, vehicle safety is an increasingly critical issue in the automobile industry. Robust and reliable vehicle detection from images acquired by a moving vehicle has numerous applications including driver assistance systems, self-guided vehicles, etc. In general, vehicle detection using adaboost classifier is very challenging due to huge within class variability. For example, vehicles may vary in shape, color, edge and size. I propose a vehicle anti-collision method exploiting multiple features base on adaboost classifier. These features include a shadow, edge, rear-lights and road characteristics. The purpose enhances the precision rate and accuracy of vehicle detection and reduces the false alarms of vehicle detection. This research technique consists of four modules: Pre-Processing, Lane Detect Processing, Vehicle Detection and Vehicle Distance Warning Processing. These modules are combining to an anti-collision system for improving safety and accident prevention.