本研究將針對使用移動攝影機偵測與追蹤移動物體的議題，進行為期兩年的 研究。主要研究項目包括物件模型的線上訓練、視覺式移動物體偵測、影像特徵偵測 與描述、以及機器人在未知動態環境中巡航等。目標是在兩年內完成以移動攝影機對 六維自由度移動物體即時偵測與追蹤之演算法，可以應用在智慧型行動機器人、輪椅 或汽車系統上，以便在未知的動態環境中具備同時定位與建圖的能力，能夠順利執行 巡航任務。 本計畫分兩年執行：在第一年，將提出基於線上訓練物件模型的移動物體偵 測演算法，以及改良影像特徵偵測、描述與比對方法，最後整合應用於智慧型行動機 器人，以便在未知的動態環境中能夠偵測、辨識與追蹤特定外型特徵的移動物體或行 人。本年度將使用擴張型卡爾曼過濾器進行系統狀態的估測，以確保移動物體訊息的 準確性。第二年將進一步使用交錯分類與切割概念進行物件與背景的分割，使發展的 演算法能夠偵測、辨識與追蹤多個不同外型特徵的移動物體或行人。此年度將改用稀 疏擴張型訊息過濾器進行系統狀態的估測，再搭配區域地圖的管理，可以規劃長距離 與大範圍的同時定位與建圖任務。本計畫將以實地測試方式評估所提相關演算法的實用性。 A two-year investigation plan is proposed in this research to focus on detection and tracking of moving object from a camera on a mobile platform. The research topics include online-training object model, moving object detection, image feature detection and description, and robot navigation in unknown and dynamic environments. The aim of this project is to develop an algorithm for detecting and tracking moving objects with six degree-of-freedom from a moving camera. The developed algorithm can provide the robot, wheelchair or car with suitable localization and mapping competence in unknown and dynamic environments in order to perform the navigation tasks. This project will be completed in two years. In the first year, a novel moving object detection algorithm based on an online-training object model is proposed. Meanwhile, a modified procedure for image feature detection, description and matching is developed for a mobile robot in order to perform detection, recognition and tracking of a moving object or pedestrian with specified characteristics. The state estimation is implemented by utilizing an extended Kalman filter (EKF) to ensure the accurate estimated information for moving object. In the second year, the concept of interleaved categorization and segmentation is applied to performed figure-background segmentation. The algorithm of moving object detection is designed to perform detection, recognition and tracking of multiple moving objects or pedestrians. The state estimation is carried out for mobile robots using sparse extended information filter (SEIF) in this year. Furthermore, being integrated with a local map management strategy, the estimation algorithm will be able to perform the tasks of simultaneous localization and mapping in long-distance and large-area environments. The developed algorithms will be integrated with real robot systems, and the integrated system will be empirically validated based on experiments to evaluate the practical usage of the proposed algorithms.
