車輛移動與交通號誌變換偵測系統

Abstract

一般而言，駕駛輔助系統在人們操作車輛時需要持續不斷地監測週遭環境，並從增進交通安全與駕駛便利性的角度提供駕駛額外的資訊。從觀察中得知駕駛在長時間等待交通號誌時經常會分心處理注意交通狀況以外的事，使得在必須起步時仍然不自覺，造成交通堵塞或被後方以喇叭提醒。本篇研究以電腦視覺為基礎提供一偵測系統，在交通號誌變換或前車前進時自動提醒駕駛。此系統包含三個模組：(1)以時間及空間記錄(spatiotemporal-profile)或掃描線(scan-line)為基礎，判斷駕駛車輛是否正在前進或停止之自身移動(ego-motion)偵測模組；(2)以Gentle AdaBoost為基礎來尋找可能存在的前方車輛，並在其移動時提醒駕駛的車輛偵測模組；(3)以顏色、形狀為基礎尋找可能的號誌，並配合背景模型在燈號轉換時發出提醒的交通號誌偵測模組。本系統希望在此較不緊急的情況下提供人們一個減輕駕駛負擔的功能。實驗結果顯示，在白天市區的場合，本系統在紅燈停止時能以95.5%的查全率(recall)與87.5%的準確率(precision)，適時提醒駕駛該再次起步以免造成交通堵塞。

In general, Driving Assistance Systems continuously monitor surrounding environment, providing information to assist human driver, in order to increase safety and convenience. By observation we found when stopping at a long-waiting traffic signal, a driver may be distracted by other tasks and not focus on traffic. When it is time to go, he or she may block the traffic or be honked by the rear car for blocking the traffic, if he or she does not move. This thesis provides a vision-based detection system which reminds the driver while it is time to start moving. The system includes three modules: (1) a spatiotemporal-profile-based or scan-line-based ego-motion detection mechanism which determines whether the driver’s vehicle is moving, (2) a Gentle AdaBoost-based vehicle detector finds possible front vehicles and sends notifications once they move, and (3) a traffic signal detector based on color/shape attributes and a background model finds possible candidates and notifies the driver once the traffic signal turns green. This system tries to provide a convenient functionality that assists people in easing driving effort in such a less critical condition. Experiments show when a driver stops at a red traffic signal in day-time urban areas, this detecting system sends notifications at a recall of 95.5% and a precision of 87.5%, therefore prevents a driver from blocking the traffic.