結合行人偵測器與物件模型演算法之視覺追蹤移動目標系統

Visual Moving Object Tracking Using Pedestrian Detector and Object Model

沈晉安(Chin-An Shen)

指導教授 ： 王銀添

淡江大學/工學院/機械與機電工程學系碩士班/碩士(2015年)

https://doi.org/10.6846/TKU.2015.00259

摘要

本論文使用物件模型結合行人偵測器，發展機器人即時偵測與追蹤移動物體系統。三個主要研究議題包括結合移動特徵偵測與行人偵測器、改良機率物件模型、與物件模型訓練調整機制等。首先，將移動特徵偵測結合行人偵測器，以限定移動物件特徵的分佈範圍。不同群的移動特徵，將被訓練成不同的移動物件；其次，修改機率物件模型的描述，確保移動物體追蹤的強健性。使用單張影像的特徵取代多張影像的特徵訓練物件模型，以提高運算效率與進行線上建模。最後，利用移動物件辨識與追蹤的回授，調適物件模型訓練的條件。調適機制分別依據定值、專家表與模糊規則等方法設計。
發展的移動物件追蹤系統進一步與視覺式同時定位與建圖系統整合，成為同時定位、建圖與移動物體追蹤系統。使用擴張型卡爾曼過濾器估測系統狀態，以及使用加速強健特徵建立視覺式環境地圖。本研究也規劃多個實驗測試範例，驗證所發展的系統之效能。

關鍵字

同時定位建圖與移動物體追蹤 ； 物件模型 ； 行人偵測器 ； 視覺式移動物體追蹤

並列摘要

This thesis presents an algorithm of robot visual moving-object tracking (MOT) based on the probabilistic object model with the pedestrian detector. Three major topics are investigated in the study including the combination of moving feature detection and pedestrian detection, the improvement of probabilistic object model, and the tuning mechanism of object model training. Firstly, the moving feature detection is integrated with a pedestrian detector to set the location boundaries of image features belonged to an object. Therefore, different groups of moving features will be trained as separated moving objects. Secondly, the representation of the probabilistic object model is modified to ensure the robustness of moving object tracking. Instead of using the image features from multiple image frames, the image features of one frame are used to train the object model for the purposes of computational efficiency as well as on-line training. Finally, the feedback of the recognition and tracking of moving objects is utilized to tune the training condition for the object model. The tuning mechanisms are designed based on fixed values, expert table, and fuzzy rules.
The developed MOT is further integrated with the visual simultaneous localization and mapping (vSLAM) to form a simultaneous localization, mapping, and moving object tracking (SLAMMOT) system. The extended Kalman filter (EKF) is used to estimate the system states and the speeded-up robust features (SURFs) are employed to represent the visual environment map. Several experiments are carried out in this research to validate the performance of the developed systems.

並列關鍵字

Visual simultaneous localization and mapping (vSLAM) ； Object model ； Pedestrian detectors ； Visual moving object tracking

參考文獻

[45] 洪敦彥，基於擴張型卡爾曼過濾器的機器人視覺式同時定位、建圖、與移動物體追蹤，淡江大學機械與機電工程學系碩士論文，2010。

[28] 陳國瑋，使用線上稀疏呈現的物件模型進行機器人並行追蹤與建圖，淡江大學機械與機電工程學系碩士論文，2012。

[43] 林冠瑜，"使用低階攝影機實現機器人視覺式SLAM"，淡江大學機械與機電工程學系碩士論文，2012。

[42] 馮盈捷，使用尺度與方向不變特徵建立機器人視覺式SLAM知稀疏與續存性地圖，淡江大學機械與機電工程學系碩士論文，2011。

[1] M.W.M.G. Dissanayake, P. Newman, S. Clark, H. Durrant-Whyte, and M. Csorba, “A Solution to the Simultaneous Localization and Map Building (SLAM) Problem,” IEEE Transactions on Robotics and Automation, vol.17, no. 3, pp. 229–241, 2001.