使用多個狀態限制卡爾曼濾波器與三焦張量幾何之視覺輔助慣性測程器

Abstract

本論文提出一套結合單一攝影機與慣性量測裝置的測程器架構。由於慣性量測裝置存在誤差累積的問題，導致單純使用慣性量測裝置的測程器在真實環境下很難得到準確的結果，必須要有其他感測器輔助。本論文以單一攝影機與慣性量測裝置進行感測器融合，由於使用了攝影機的資訊，慣性量測裝置的誤差可以被有效地抑制住，使得慣性量測裝置可以提供實際地圖尺度。 本論文使用在三張影像中存在的三焦張量幾何關係作為攝影機所提供的量測資訊，此方法可以不需要估測特徵點在空間中的位置，也就是不需要進行環境的重建。同時本論文將三張影像分別對應到的攝影機姿態在濾波器中修正，形成一多個狀態限制的卡爾曼濾波器，以此提出一個在計算量與精確度間取得平衡的滑動視窗式測程法，相較於現存視覺式測程法或是同時定位與地圖建立的方法，本論文提出的架構更符合自我軌跡估測的測程需求並且適用於即時導航系統。本論文同時提出基於三視角幾何的隨機抽樣一致演算法，它可以有效地排除比對錯誤或是落於移動物體上的特徵點，使得整體演算法在動態的環境下仍有可靠的結果。最後，本論文實作一套整合單一攝影機、慣性量測裝置以及GPS的硬體設備並以在真實環境下的資料驗證所提出的演算法。

This thesis presents an odometer architecture which combines a monocular camera and an inertial measurement unit (IMU). Due to error accumulation problem, it is very hard to get reliable result by only using the IMU, therefore there is a need to fuse the information with other sensors. In this dissertation, a monocular camera and an IMU are used for sensor fusion. By using the information of the camera, the error of IMU can be effectively constrained, so that the IMU can provide real scale. In this dissertation, the trifocal tensor geometry relationship between three images is used as camera measurement information, which makes the proposed method without estimating the 3D position of feature point. In other words, the proposed method does not have to reconstruct environment. Meanwhile, the camera pose corresponding to each of the three images are refined in filter to form a multi-state constraint Kalman filter. Consequently, this dissertation proposes a sliding window odometry which has a balance between computational cost and accuracy. Compared with traditional visual odometry or simultaneous localization and mapping (SLAM) method, the proposed method not only meets the requirement of odometer in ego-motion estimation, but also suit for real-time application. This dissertation further proposes a random sample consensus (RANSAC) algorithm which is based on three views geometry. The RANSAC algorithm can effectively reject feature points which are mismatch or located on independently moving objects, thus it make the overall algorithm capable of operating in dynamic environment. Finally, a hardware which integrates with a monocular camera, an IMU and a GPS is implemented, and experiments are conducted to validate the proposed method in real environment.