人物定位演算法之實際應用

Abstract

隨著自動監控系統的普及，對於即時人物定位的需求也愈來愈多，如何建立有效且低成本的人物定位系統已是近年來重要的議題之一。在之前的系統中，都是使用四支攝影機來架設系統，然而，四支攝影機的架設較為困難、硬體成本高且定位時間也相對較長。因此，在本論文中，我們試著在同系統上減少一支甚至兩支攝影機。首先，根據場景的特性，我們從三種前景偵測方法中選出其中一種做為即時人物定位的方法。接著描述在之前的定位方法下，減少攝影機會產生什麼問題，並找出最適合三支攝影機的擺放位置。最後，我們提出了數個改進方法讓系統更穩定。實驗結果顯示，即使只剩下三台攝影機，系統仍然保持著一定程度的定位水準。

With the popularity of automatic surveillance system, demand for real-time people localization system grows fast. Thus, constructing the system with high efficiency and low cost has become an important issue in recent years. In our previous work, four cameras are used to build the system. However, the four-camera system is complex to calibrate and has high cost in hardware and computation time. Thus, we try to build similar system using less cameras. First, one foreground detection methods out of three is adopted to a real-time localization system based on the characteristic in the selected scene. Second, problems of directly reducing cameras of the previously established system are described, and two better configurations of three-camera system is proposed. Finally, we also propose possible improvements to the robustness of the system. Experimental results shows the system can perform satisfactorily with just three cameras.