藉由模擬行人行為之不尋常行為偵測

Abstract

在本論文中，我們提出一套自動化的行人異常行為偵測之監控系統。不同於一般監控系統是以環境場景分析的觀點出發來偵測異常行為的發生，我們是採取以行人的觀點為我們的出發點。行人的行為有千奇百種，若要仔細的去一一描述各種不同的行為是有困難的。但在一般的情況下，行人的行為是有跡可循的，並非如一般想像的隨機變動；亦即行人的行為都會因當時的周遭環境而採取在當下最適當的應對行為。因此描述周遭環境因素對行人行為所造成影響，將有助於我們去模擬各種行人的行為。在此，我們藉由行人的特性來描述行人的行為與周遭環境因素之間的關係，這些行人的特性包含了：目的地、舒適距離和朋友關係等。之後再去描述這些不同的特性對於行為所造成影響，來幫助我們來預測行人可能的行為模式進而幫助我們達到偵測異常行為發生的目的。相較於去建立所謂的場景模型，透過以行人的觀點來偵測異常行為的發生，我們將可以偵測出更多種抽象的異常行為發生。

In this paper, we propose a surveillance system which is able to detect abnormal behaviors automatically. Unlike other approaches, we do not detect abnormal behaviors by constructing a scene model with normal behaviors. Instead, we emphasize the important aspect of pedestrian since it is the pedestrian who performs abnormal behaviors. There are actually various pedestrian behaviors and it is difficult to express every pedestrian behavior in detail. However, it is still possible to model pedestrian behavior since pedestrian behavior, in general, always reflects some kinds of close relationship with the current environment. Hence, we first take some social factors of the environment into account for the estimation of pedestrian properties. After that, we predict the possible pedestrian behaviors by using these estimated pedestrian properties. An abnormal behavior can be detected as long as the behavior of pedestrian is very different from the predicted behaviors. With this system, we are capable of detecting more abstract abnormal behaviors than before.