- 學位論文
卡爾曼濾波器之視覺慣性目標姿態估測與感測器間自我校正
Kalman Filter Based Visual Inertial Target Pose Estimation with Sensor-to-sensor Self Calibration
摘要
慣性感測器與視覺感測器的組合常用於姿態估測以及機器人導航，在本研究中，為了解決機器人自主對接以及充電問題，提出了一套以擴展型卡爾曼濾波器為基底之慣性-視覺目標六維姿態估測器，透過計算模型亞可比陣列可將精準度高但非線性量測模型應用於線性的卡爾曼濾波器，結合複數個影像針對同一特徵之幾何限制以及待測目標之六維姿態作為量測模型，此方法的優點為系統狀態不需要估測三維特徵位置。換句話說，這個方法不需要重建環境，大幅度地減少演算法複雜度，更符合即時運算的構想。要如何正確地結合兩種感測器資訊以及提高整體估測器的精準度，兩感測器之間的六維相對姿態精準度佔有極為重要的角色，錯誤的校正過程會導致偏移誤差並且降低估測的精準度，甚至導致估測器量測發散，本研究為使用線上運算的概念即時校正感測器之間的六維姿態，在目標估測的平均誤差可達到2.754公分以及0.702度。最後由實際運行結果顯示結合慣性視覺感測器之目標六維姿態估測器的精準度，並且分析有無感測器自主校正在目標估測上的影響，不僅僅能提升在於快速運動下的目標估測精準度，也同時能夠正確的校正感測器之間的相對姿態。
並列摘要
Inertial and visual sensors are usually used in pose estimation and robot navigation. This research presents an Extended Kalman filter (EKF) based visual-inertial target 6-DoF pose estimator for robot autonomous docking and recharging problems. The nonlinear measurement model which is more accurate but highly complex can be applied to linear Kalman filter through calculating Jacobian without losing its accuracy. Combining geometry constraints of the mulit-camera views and the target 6-DoF pose are served as the measurement model. This model does not require including 3D feature position in the state vector. In other words, this method doesn’t need to reconstruct the environment which can reduce the algorithm complexity and make it more effective for real-time executing. Correct data fusion and hence overall estimator accuracy, depends on the accurate calibration of the 6-DoF relative pose between two sensors. The errors in the IMU-camera extrinsic calibration process cause biases that reduce the estimation accuracy and can even lead to divergence of any estimator processing the measurements from both sensors. This research, we describe an algorithm which can applies the concept of online calibrating sensors relative pose. The average error in the estimation result is 2.754cm and 0.702 degree. The experimental results are presented that the accuracy of visual-inertial target pose estimator and analysis the impact of the online sensor-to-sensor self-calibration. This algorithm can not only enhance the estimation accuracy during high speed movement, but also calibrating sensor-to-sensor relative pose accurately.