多台攝影機監視系統下的物體追蹤與對應

Abstract

在本文中，我們提出了一個基於三維粒子濾波器的多物體追蹤及對應技術，利用多個二維物體追蹤的結果，建立起物體在三維空間中的機率分佈，並透過二維三維資訊的交換，更新及預測這個分佈的變化。同時，這個三維的機率分佈，也可以幫助我們修正二維追蹤的結果，以及建立起各視野間物體的對應關係，並且能夠自動的修正錯誤的對應關係。有了物體的對應關係，就可以解決物體交錯所造成的追蹤錯誤，並且能夠更加正確的估測三維物體的機率分佈，使多物體的追蹤和對應更加準確而可靠。

In this thesis, we propose a 3-D particle filter based objects tracking and correspondence system. Through the 2-D tracking results, we can predict and update the probability distribution of moving objects in 3-D domain. With such probability, we can not only refine the 2-D tracking results but also construct the correspondence of objects in different camera views. In addition, our system is able to correct the correspondence automatically based on some 2-D and 3-D clues. With the object correspondence, the occlusion problem can be solved easily and the 3-D probability distribution of moving objects can also be estimated more precisely. These advantages make the results of objects tracking and correspondence more robust and reliable.