物件追蹤為電腦視覺領域重要的議題，可應用於監控系統與人機介面。如何準確估測物件大小並採用合適的物件特徵，以增進追蹤準確度，實為重要的課題。本論文以人類頭部為追蹤目標，以粒子濾波法為基礎，建立適用於非線性與非高斯機率描述的機率狀態轉換與量測的系統。我們將偵測機制整合使追蹤系統，並提出在追蹤的過程中，依據追蹤結果與目標物件的顏色相似度，啟動以不同特徵為基礎的頭部定位系統之方案，重置追蹤系統的目標物件顏色資訊和目前畫面的頭部大小。實驗結果顯示，當人頭部隨意運動，快速移動和對攝影機有距離遠近改變時，本系統仍可達成不錯的追蹤準確性。 Object tracking is an important technique in computer vision, and it can be applied in applications such as visual surveillance and human-robot interaction. How to estimate object scale accurately and choose proper feature to improve tracking accuracy is an important issue. In this paper, our tracking system tracks human heads with particle filter with non-linear and non-Gaussian state transition and measurement. We integrate head detection into tracking system and propose to start head localization with various features based on color similarity of tracking measurement. We reset target color histogram and head scale if needed. Experimental results show that our head tracking system has good tracking accuracy under human regular motion, fast motion and distance variation between the target and the camera.
