智慧型社區安全影像監控系統之研究

Abstract

近年來，智慧型社區安全越來越受到國人所重視，因此在本論文中我們提出針對智慧型影像社區安全系統有關的三個子系統: 第一部分利用Boosted訓練一高準確度之車輛偵測系統; 第二部分使用以人形外型樣板為基礎的Codebook來做為人形偵測系統; 第三部分使用AdaBoost為主的人臉偵測系統。Boosted方面是使用AdaBoost 訓練可適用於白天及傍晚的車輛偵測器，此偵測器擁有高偵測率，但誤判率也相對比較高，被誤判的內容也相對複雜，因此我們建立了一套可針對白天以及傍晚兩種情況濾除誤判的系統，白天使用邊緣複雜度來做為白天誤報減少，傍晚使用亮度複雜度之直方圖來做為傍晚誤報減少，在不同的環境下會自動切換演算法來達到車輛偵測系統之穩定。在人形偵測方面，使用背景相減法(Background Subtraction)來做為取出移動物體，並將移動物體搭配簡單的軌跡與狀態判別模式，為之後人形辨識部分上提供一些必要資訊，以及降低誤判(False-Alarm)，並以人形樣版為基礎的Codebook來實現人形偵測系統。確定為人形時，接著使用2D Haar為特徵以AdaBoost 學習演算法來進行人臉偵測，搭配直方圖均化，來提人臉偵測率。從我們的實驗結果可以看出我們所提出的方法使用在社區安全監控系統是有不錯的表現。

In recent years, much attention has been directed towards intelligent community security systems. In this study, we propose three sub-systems for intelligent community security systems. The first sub-system uses boosting to train a high-accuracy vehicle detection system. In the second sub-system, we choose a shape-based method using a codebook to classify humans from other objects. The final sub-system uses AdaBoost to detect faces. We use the AdaBoost algorithm to train a vehicle classifier that can operate during the day and evening. The classifier has a high detection rate, but the false alarm rate is also relatively high. Moreover, the content of the false alarms is complicated. Therefore, we have also developed two algorithms to reduce the false alarm rate, one for daytime using edge complexity and another for evening using histogram matching and an intensity mask. In the different environments, an automatic switch algorithm is used to achieve the required stability for the vehicle detection system. We use background subtraction to segment a moving object and simple trajectory tracking and condition judgment to provide data for the human detection algorithm and decrease the false alarm rate. We use a shape-based method using a codebook to implement the human detection system. If the moving object is human, we use 2D-Haar features to train an AdaBoost algorithm for face detection and use a histogram specification to improve the results. Finally, our experimental results show that our proposed approaches perform well for community security image surveillance systems.