應用於多樓層停車場之視覺輔助慣性定位系統

Abstract

近年來應用全球衛星定位系統(GPS)定位導航的應用不勝枚舉，然而當衛星訊號受到遮蔽例如進入建築物內時，全球衛星定位系統將無法提供連續、準確的定位導航服務。本論文設計的系統利用慣性量測元件與相機兩不同特性之元件做為感測器，以卡爾曼濾波器以及種種元件互補的機制，使車輛在沒有全球衛星定位系統訊號的室內多樓層停車場擁有自我定位的能力，以期車輛得以在停車場內以有效率的方式尋找車位。本論文的主要貢獻有四，第一，本論文提出的資料融合架構，利用路況判別等機制使得各元件能發揮其元件特性，在適當的時機將可靠的資料交給系統，並非一味的直接將兩感測器的資料作融合，而不考慮資料當時的可靠程度。第二，本論文的實驗採用一般車輛作為載具，而並非使用可控制方向和速度的自走式機器人，大大增加了系統的實用性。第三，本系統經過日間及夜間不同環境的測試，也經過不同廠牌車輛的測試，而實驗中也遭遇行人等移動物體干擾和路面起伏不定等狀況，確保系統的穩健性。第四，本系統設計的電腦視覺部分每秒只需一張畫面做計算，相較於其他研究，本論文提出之系統可以大幅降低硬體之需求。經實驗證明，本論文之視覺輔助慣性定位系統大幅降低純慣性定位系統之誤差，本系統於停車場對車輛的軌跡估測之精準度已足夠達成車輛於停車場中自我定位用以尋找車位之目的。

In recent years, global positioning system (GPS) have been widely used in localization and navigation application, however, GPS can’t provide continuous and accurate service when the GPS signals are not available (e.g., indoors, underground, etc.). In this paper, we present an approach to fuse measurements from an inertial measurement unit (IMU) and a camera for self-localization in GPS-denied multi-storey parking lot. The contributions of our work are fourfold. First, the introduced fuse architecture, instead of fusing the IMU information and the vision directly, we incorporate the IMU information and the vision in different ways when systems encounter different condition. Second, we choose cars as the vehicles, instead of robots, which are controllable in orientation and velocity, this condition make this system more practical. Third, the system in this work has been testing in actual parking structure, the test data obtained from different brand of cars during daytime and nighttime. Fourth, the vision system used in this work uses a slow-frame-rate monocular camera, the proposed system can significantly reduce the demand for hardware. The experiments shows that the vision-aided inertial localization system significantly reduce the error of inertial localization system, and the accuracy of the vision-aided inertial localization system is sufficient to achieve self-positioning of vehicles in the parking lot for the purpose of looking for parking spaces.