- 學位論文
人體偵測與追蹤演算法的分類與評估
A Taxonomy and Evaluation of Human Detection and Tracking Algorithms
摘要
人體偵測與追蹤為電腦視覺中一項重要的研究主題，已被廣泛利用到監視系統、 人機介面、行為分析等應用上面。現今有許多的人體偵測與追蹤系統，當面對偵測環 境改變時，都會導致偵測效果不佳，並且沒有一套整合的平台來評估不同的人體偵測 與追蹤系統。本論文將人體偵測與追蹤演算法區分成以下主要步驟： (1)影像擷取，以影像擷取裝置得到影像序列； (2)前景偵測，使用快速且正確的方式將物體從背景中分離； (3)人體偵測，在前景偵測結果中以人體特徵來找出人體位置； (4)人體追蹤，在影像序列中找出人體移動的軌跡。 利用上述分類，我們可以探討這些步驟的設計需求，完成這些步驟的各種演算法 的優、缺點，並實驗如何整合出一個最佳的人體偵測與追蹤系統。 為了達到整合各種不同演算法的目的，本篇論文在MATLAB環境下設計一套人體 偵測與追蹤的整合測試平台，將人體偵測與追蹤系統的三個主要部份：前景偵測、人 體偵測、與人體追蹤，每個主要部分的程式都模組化，使得系統可以在各個模組彈性 的測試不同的演算法。我們亦提出了幾組公開測試資料及效能評估計算方法，便於比 較不同人體偵測與追蹤演算法整合後的效能。 利用本論文所提出的整合測試平台，我們實驗出一組最有效率且準確的人體偵測 與追蹤系統，另外我們也提出了一些新技術，修改了現有的演算法使得整合後的系統 更為快速及準確。 關鍵詞：前景偵測、人體偵測、人體追蹤、多重假設追蹤
並列摘要
Human detection and tracking is an important research in computer vision, and it has been extensively applied to the surveillance systems, the human-machine interfaces, behavior analysis, and so on. Nowadays, there are many human detection and tracking systems. They will produce false detection while they are confronted with the changing environment, such as lighting changes. There is no any integrated platform to evaluate the performance of these different human detection and tracking systems. We have proposed a framework that divides the human detection and tracking algorithms into four major components as follows: (1) Image acquisition: get the image sequences with the image acquisition device; (2) Foreground detection: use fast and correct methods to separate the object from the background; (3) Human detection: find out the human position using the human characteristics from the result of foreground detection; (4) Human tracking: find out the tracks of humans in the image sequences. According to the foregoing taxonomy, we can define the requirements of these components and then analyze the advantages and drawbacks of various algorithms. Through experiments, it is possible to integrate an optimum system for human detection and tracking. In order to allow flexible integration of different algorithms for these components, we use MATLAB to design a test platform of human detection and tracking system, which is divided into three major components: foreground detection, human detection and human tracking. Each component has a specific program module. These program modules can be replaced to test different algorithms. We have also proposed several public test data and measurement that allow fair comparison of different human detection and tracking systems after integrating different algorithms. By using the proposed integrated test platform, we have designed an efficient and accurate human detection and tracking system. Moreover, some new methods are proposed to modify existing algorithms to further improve the performance. Keyword：Foreground Detection、Human Detection、Human Tracking