- 學位論文
基於單眼視覺里程計與路面標誌疊合之結構路面車輛 定位系統
Vehicle Localization System based on Monocular Visual Odometry and Road Mark Matching on Structured Road
摘要
對於自動駕駛車來說，自駕車定位是一件重要的事，有許多不同的感測器可以用來定位，包括全球定位系統、慣性感測器、光達、相機等，不過全球定位系統易受到外在環境影響造成結果不精確，光達較為精確可是昂貴，慣性感測器用久後會有累積誤差的問題，此時相機是較好的選擇，視覺里程計便是使用影像資訊求得相機相對姿態，不過就算是利用影像定位還是有累積誤差，地圖上的路面標誌資訊可以協助修正位置。 在這篇論文中使用單眼視覺里程計，不過單眼視覺里程計不能求出相機絕對的位移，這部分利用全球定位系統或慣性感測器的資訊彌補，在視覺里程計中，車輛的運動學模型納入考量，並求得車輛在平面上的旋轉，為了減少累積誤差的影響，運動估測是從所有序列的影像對關鍵序列的影像做計算，關鍵序列會隨著時間更新。 在路面標誌方面，路面標誌偵測基於機器學習方法被提出，不過計算時間有點長，所以我們提出路標追蹤基於前一刻的路標資訊，只需要路面標誌偵測八分之一的時間，為了避免不必要的路標偵測，路標偵測會根據車輛在地圖上的位置判斷是否需要運作。 為了將兩種資訊結合，此篇論文使用卡爾曼濾波器，不過動態模型是非線性方程式，無跡轉換來處理高斯分布在非線性方程式上的轉換，根據實驗結果，雖然視覺里程計會因為位移資訊的不準確造成結果不精確，不過藉由路面標誌的資訊可以修正回來，路標偵測在無陰影的情況下可以找到路標。
並列摘要
With the development of autonomous vehicle, vehicle localization becomes more and more significant. Localization can be achieved by many different sensors, including LIDAR, GPS, IMU and camera. However, GPS is inaccurate when GPS signal is blocked by external environment. LIDAR is too expensive and cumulative error exist from IMU. Camera is appropriate for localization. To estimate relative pose of camera, visual odometry is used. But there is cumulative error in visual odometry after using for a while. To obtain a more accurate localization result, a global map with recorded road marks can be used. In visual odometry part, monocular camera is used in research. However, translation movement scale cannot be obtained by monocular visual odometry. Translation is obtained from GPS or IMU in research. Kinematic model for ground vehicle is used to obtain planner motion. To reduce cumulative error, every frame is used to estimate motion relative to keyframe. In localization based on road mark part, road mark detection based on machine learning is presented. But it takes a lot of time to do road mark detection. Road mark tracking based on prior information is also present. Computational time in road mark tracking is one eighth of computational time in road mark detection. To avoid unnecessary road mark detection, road mark detection is active or not based on car position on map. If car is nearby road mark, road mark detection is active. To combine car position information from two source, Kalman filter is used. But motion model is nonlinear. Unscented Transform is used to solve nonlinear transformation of Gaussian distribution. The experiment shows that result from visual odometry sometimes is not precise due to inaccurate translation movement. But result can be fixed by measurement from detected road mark in Unscented Kalman filter. Road mark detection can find road mark in consistent light condition.