- 學位論文
融合影像與圖資之自駕運具定位系統性能提升研究
Performance Enhancement of Localization System for autonomous Vehicle with Computer Vision and Map-based Sensor Fusion
摘要
本論文致力於研發使用中/低價位的感知器透過融合技術來達到SAE針對自駕載具分級之LV3所需之定位系統，同時，本論文所定義之準確度乃以整車系統在穩定且安全的運作情形為前提下，定位系統資訊充分且足夠準確，因此並不需要如同SLAM追求數公分的誤差而耗費大量成本，使定位系統性價比低且受限於中低速場域。 為了使用中低價位之感知器來實現全自動駕駛載具所需之定位系統，本論文首先根據交通部之道路法規以及載具規格進行準確度的量化定義，而後提出一套以圖資導引技術為基礎，透過動態拘束型無損卡爾曼濾波器(DCUKF)融合影像辨識模組所提供之區域環境資訊(側向車道線或縱向停止線)、衛星定位的絕對位置資訊與慣性感測的車輛運動資訊，動態調整取樣點分布之限制邊界，並搭配錯誤偵測模組監測各個感知器的當前狀態，適時調整DCUKF所使用之量測協方差矩陣與切換量測更新階段所使用的感知器資訊，提升定位系統抗雜訊干擾之能力並獲得最佳估測結果。 藉由上述方法，可使自動駕駛系統於使用中、低價位感知器的情況下，仍擁有足夠準度的定位功能，並於衛星定位系統或影像辨識系統之單一感知器失效時，延長可信的定位估測結果，給予自動駕駛系統較安全的行駛行為，實現低價位且兼具準度與穩定安全之自動駕駛載具定位系統。 最後，本論文於工業技術研究院中興院區進行定位系統之實驗數據採集，並分別進行可行性測試、演算法狀態分析與穩定性測試三種實驗，測試出定位系統準確度確實符合目標之公分等級定位，且定位系統在長時間運作時的狀態皆無發散震盪的情況發生，同時錯誤偵測亦能正確指示出當前演算法狀況。
並列摘要
This thesis proposes a low-cost lane-level localization algorithm based on the concept of map guidance for autonomous vehicle which level of self-driving is above level three under SAE International's J3016 standard. The localization system could be used in specific closed field, such as campus, terminals. The lane-level localization means that locating the vehicle with centimeter-level accuracy. In order to achieve the requirement of centimeter accuracy with a mid/low-cost GNSS, this thesis proposed a Dynamic-constrained Unscented Kalman filter (DCUKF) fusing computer vision, low-cost GNSS, and inertial navigation system as measurement inputs. The DCUKF is an observer of the vehicle pose (position and heading), and it is based on Unscented Kalman filter (UKF) which uses the the sigma points around mean state in the previous time step to estimate the current mean value of state and its covariance. For the estimation result, digital maps will provide lane boundary and attributes which is used as a global constraint and lanemarks outputed from vision identification module will be used as a local constraint. By utilizing different constraints on sigma point of UKF, the sigma points can be realigned according to the position reference, increasing the precision of vehicle position and enhance the ability of bias/noise resistance for localization system. In addition, fault detection module will monitor the sensor status and estimated results to prevent the motion controller from using the error estimated results causing a safety issue.