即時多物件追蹤與分析系統

Abstract

由於軍事和社會安全的需要，影像分析和物件追蹤在近年來逐漸成為一個被熱門研究的領域。影像分析和物件追蹤主要是由前端的前景擷取和後端的物件分析這兩個過程所組成，前端的前景擷取效果嚴重影響著之後物件分析的難度和分析準確性，因此本論文對於系統中之背景與前景擷取之效率與正確性特別加以探討與研究。這篇論文先藉由實作許多常見且直觀的背景產生和前景擷取演算法以分析其優缺點，並針對即時性和節省記憶體使用量方面提出了新的背景產生方法。在前景擷取過程中，並針對常導致物件分析出錯的影子利用Sobel Filter來達到預先消除的目的。 在物件追蹤方面，本篇論文採用Kalman filter為核心來對物件運行的軌跡和物件大小變化來做猜測和修正。此外，針對複數物件之間可能會發生的重疊合併與分離做出辨識並持續追蹤。此系統也可以依使用目的讓使用者自由定義所需辨識之物件的特徵、所須監視的區域與所需發出警報的事件。 此系統經過了多部影片的測試，包括室內、室外、人車混合、高速公路與高雜訊環境下之測試，並實際與架設在台灣大學育成中心內之一監視器結合。經過長時間的測試結果，顯示出此系統在物件辨識成功率與即時性上面都有著不錯的成效，並留有許多擴充的空間。

As a result of the need of military and social security, image analysis and object tracking has become the field studied popularly in the recent years. Image analysis and object tracking consist of two main steps ‘foreground extraction’ and ‘object analysis and tracking’. The result of foreground extraction has the significant influence on the accuracy of object analysis and tracking, so this paper emphasizes the efficiency and accuracy of background generation and foreground extraction. Firstly, this paper implements many simple and popular background generation and foreground extraction algorithms, then analysis the advantage and disadvantage. Secondly, this paper also presents a new method of background generations that fulfill the requirement of real-time and reduce the usage of memory. For foreground extraction, this system detects and removes shadow by sobel filter. Object tracking is performed by applying a kalman filter to predict and correct the trajectory and size change of object. In addition, tracking and recognition of merge and split of multiple objects are also available. This system may depend on the user’s goal and let user define freely the feature of object which would be tracked, the region would be monitored and the activity would let the system make an alert. This system passed through many testing videos, including in-door, out-door, mix of human and vehicle, the highway and the high-noise environment. In addition, to test and evaluate the algorithm and method, this work deploys the system in a real living environment out of our lab. The experiment result present that this system have good efficiency in real-time and high accuracy, it is also flexible and robust.