| 摘要: | 本論文以擴張型卡爾曼過濾器(extended Kalman filter, EKF)建立視覺式同時定位、建圖、與移動物體追蹤(simultaneous localization, mapping and moving object tracking, SLAMMOT)系統。此SLAMMOT系統的感測器使用視覺系統為唯一的感測裝置，搭配加速強健特徵(Speeded-Up Robust Features, SURF)的偵測方法進行環境中影像特徵的偵測，並且依據影像訊息求算特徵在空間中的三維座標，以建立SURF特徵式地圖。在估測器方面，本研究使用EKF方法遞廻估測機器人與影像特徵的狀態，並且規劃符合SURF特徵式地圖的資料新增、刪除、與更新等程序。移動物體運動模型方面，本系統使用交互式多模型(interacting multiple model, IMM)估測器對移動物體進行追蹤任務，使系統能同時處理靜態與動態移動的物體。本論文針對以上SLAMMOT系統三個部份所需的理論與技術進行探討，並且對文獻現有方法提出改善的方案。
本論文也以模擬與實測方式，驗證所提方案的可行性。首先，使用單眼視覺實現具備未知輸入之系統的同時自我定位與特徵式地圖建立，同時也解決單眼視覺的影像深度量測問題。其次，以IMM估測器實現視覺式移動物體的追蹤，使用雙眼視覺做為感測器以簡化移動物體三維座標量測的問題。最後，整合EKF估測方法、特徵式地圖、與IMM估測器，實現雙眼視覺式同時定位、建圖、與移動物體追蹤之任務。
In this thesis, the visual simultaneous localization, mapping and moving object tracking (SLAMMOT) is established by using the extended Kalman filter (EKF). The theory and methodology of vision sensing, state estimation and motion modeling of the SLAMMOT system will be investigated in this thesis. For sensor perception, the visual system is the only sensing device in the SLAMMOT system. Meanwhile, the method of detecting the speeded-up robust features (SURF) is utilized to detect the image features in the environment. According to the extracted data of the image features, three-dimensional coordinates of the features are calculated and then the feature-based map based on SURF are built. For state estimation, the EKF is employed to predict and update the states of the robot and the features recursively. Furthermore, the procedures of adding, erasing and updating the data of the SURF in the map are planned. In modeling the motion of moving objects, the interacting multiple model (IMM) estimator is utilized to track the moving objects. Therefore, the system can handle both the stationary and moving objects at the same time.
The proposed algorithms are validated through computer simulations and experimental works on real systems. First, simultaneous localization and feature-based mapping are implemented on a free-moving monocular vision system with unknown inputs. Meanwhile, the problem of determining image-depth in monocular vision is solved. Second, the vision-based moving object tracking is performed by using the IMM estimator. Furthermore, the sensor is replaced by a binocular vision to simplify the problem of calculating the three-dimensional coordinate of moving objects. Finally, the tasks simultaneous localization, mapping and moving object tracking using a binocular vision are implemented by integrating the EKF method, feature-based map and IMM estimator. |
