| 摘要: | 典型的視訊監控目的，是將影像中感興趣的物件擷取出來作分析、辨識等應用。而物件擷取技術則通常建立在背景移除法的基礎上。本論文以一個公開的視訊資料庫來評估7種主要的背景移除法：FD、AMF、Stauffer GMM、Zivkovic GMM、KDE、Eigenbackground、Codebook，我們採用ROC曲線作為背景移除性能評估方法。有別於前人在背景移除法領域中所做的檢視(survey)項目，本論文特別針對了這些演算法在嵌入式硬體化實現的適合性進行全面探討，我們比較每個演算法的執行效能(速度)、記憶體使用量、程式碼大小，以利作為後續從事背景移除法的嵌入式硬體實作之重要參考。從實驗的結果顯示出，近似中值濾波法幾乎在各種影片分類中，準確度與精確度的表現皆為前兩名，再加上其演算法簡單、記憶體使用量較少，對於移動物件偵測的嵌入式系統設計是個不錯的選擇。此外這樣的成果也間接證實了背景移除法的難易程度與準確度及精確度的高低並沒有關聯性。
;Video surveillance is typically used to capture moving subjects which users may be interested in from an image sequence, then applying it into different software applications for analysis and identification afterwards. These kinds of techniques are usually based on background subtraction (BS). In this paper, seven popular BS algorithms (FD, AMF, Stauffer GMM, Zivkovic GMM, KDE, Eigenbackground, Codebook) are compared and evaluated with open source video database to rate their performance. We adopted the receiver operating characteristic (ROC) curve as the evaluation metric to compare the result of BS algorithm, foreground object, with accurate ground truth data. In contrast with previous BS studies, this paper is especially focused on the complexity of implementing these methods on hardware device, like FPGA. Several properties of each algorithm will also be discussed in the article including accuracy, precision, efficiency, memory usage, and code size. The findings will provide reference for future BS algorithm applications on embedded hardware systems. Our results show that approximated median filtering is precise and performs superior in every evaluative category. Considering its ease of use and minimal memory requirements, it is a pragmatic choice for embedded systems design. Furthermore, our findings reveal no significant difference between accuracy of the results from the various BS methods used. |
