非重疊攝影機間之機率式交通流量建模方法

Abstract

在延伸現有視覺監視系統至全面性交通監控的過程中，如何推斷多 攝影機間之交通狀況是一項非常重要的課題。在本論文中，我們提出 一個於非重疊攝影機間之機率式交通流量建模的有效方法。基於物體於攝影機間移動時其移動時間會符合某總體模型之假設，並藉由連續估計此模型之參數，我們可以推斷出在不可視區域中之動態交通狀況。原則上如果我們知道攝影機間之移動物體的對應關係，則移動時間的總體模型的參數即可被估計出來。然而尋找物體間之對應關係在電腦視覺領域中仍然是一個懸而未決的問題。在本文中，我們將非重疊攝影機間物體對應關係之建立與總體移動時間模型之參數估計視為一個統整性的最佳化問題，並利用期望值最大化演算法來反覆尋找最佳的物體對應關係以及總體模型之參數解。在實際場景中的實驗結果證明了我們的方法能有效地估計出總體移動時間模型之參數並準確推斷出動態的交通情況。

The ability to infer the traffic status across multiple cameras allows the extended use of existing vision-based surveillance systems to global traffic monitoring. In this paper, we propose an efficient algorithm to probabilistically model the dynamic traffic flow between non-overlapping FOVs. By assuming the transition time of object moving across cameras follows a global model and consecutively estimate the model parameters, we may infer the time-varying traffic status in the unseen region. In principle, the parameters of the transition time model can be estimated if the object correspondence between non-overlapping FOVs is known. However, object correspondence itself is still an unsolved problem in computer vision. In this paper, we model object correspondence and the parameters estimation as a unified problem in a proposed Expectation-Maximization (EM) based framework. By treating object correspondence as a latent random variable, the proposed framework can iteratively search for the optimal object correspondence and model parameters. Experimental results on real data show the accuracy of dynamic model estimation and the beneficial inference of the traffic status.