以膚色資訊加速之AdaBoost即時人臉偵測系統

Abstract

治安問題一直是政府及社會大眾常討論的課題之一，尤以層出不窮的竊盜案最為人所詬病。本篇論文提出了一套即時人臉偵測系統以輔助監控系統，僅需要一般的攝影器材就能達到偵測及追蹤人臉的功能。本系統由六個模組所組成，其中包含了膚色分割、候選臉部區域篩選、特徵訓練、視窗掃描、視窗偵測及臉部視窗膚色驗證。由於影像中大部分面積為非膚色，而人臉上佈滿膚色資訊，故本實驗充分利用膚色資訊作為前處理及人臉驗證，可大幅降低系統運算量。在偵測方面，使用AdaBoost演算法挑選出最佳的矩型特徵(Rectangle feature)組合，並利用積分影像(Integral image)可使掃描視窗快速計算出矩型特徵值。經實驗證明，本研究之偵測人臉方法能達到96.12%偵測率，且執行時間為OpenCV的12%~96%，在640x480的影像中，平均執行時間僅需80毫秒。

Law and order problem has been one of the topics that often discussed by the government and the society. Especially in the theft cases most criticized. This thesis proposes a set of real-time face detection system to supplement the monitoring system. The system includes six functions, Skin Color Segmentation, Filter Candidate Face Region, Training Features, Scanning Windows, Detection Window, Verify Face Window. As the image, non-skin color in most area, and the person’s face covered with skin color information. So, full use of skin color information as pre-process and verify faces can be greatly reduced system operator. In the detection, the use of AdaBoost algorithm to select the best combination of rectangle features. And the use of integral image, the scan windows to quickly calculate the value of rectangle feature. From experimental result, it was found 96.12 percent of correct face detection is achieved using the proposed method, and the execution time of 0.12~0.96 times the OpenCV. In the 640x480 image, the average execution time of only 80 ms.