比較Real和Gentle Adaboost應用於人臉偵測

Abstract

臉偵測的技術應用於非常廣泛的領域中，其中最常見到的是在相機這種小型的嵌入式平台上，但嵌入式平台資源拮据，使用快速且簡單的方法視必成為重點，因此本篇論文中探討利用Adaboost快速計算的特性，並比較傳統與改良過後的Adaboost演算法應用於人臉偵測上的效果，最後找出效能較高的一種。 首先會介紹 DaVinci 平台的特性及所使用的工具，接著說明傳統的Adaboost演算法，其應用於人臉偵測上相當的快速也有不錯的效果，但演算法存在一些缺點，這些問題能夠加以改進絕對能夠對速度及效能更加的提昇。而改進的方法又分為Real和Gentle兩種，我們會從所改良部分、選出的特徵、實作效能等方面去做比較，觀察兩種方法模擬出的數據，最後移植於 DaVinci 平台作即時影像的偵測。在 DaVinci 中，ARM會透過與 DSP 之間的 Share memory 將影像傳給 DSP 做人臉偵測，首先會將影像切割成多個檢測視窗，再運用積分圖計算特徵值，當此視窗通過所有 stage 檢測為人臉時，還會做膚色區域檢查，確定否是為人臉，最後在影像中框出此視窗範圍。DSP端需要75ms處理完一張影像，而每秒可輸出十三張左右影像。

Face detection is wildly used in lots of fields. This technique is commonly applied in the small embedded system such as camera. Because of the poorness of embedded system, faster and simpler detective method is important. This thesis is focused on fast computing through Adaboost. We will compare the performance of traditional and improved Adaboost for the face detection and find the more powerful one. First, we introduce the characteristic of the DaVinci embedded system and the tool. Then we introduce the traditional Adaboost algorithm which has fast and good performance for the face detection. But the algorithm still has some problems that can be improved to boost speed and performance. The improvement can be classified two different methods Real and Gentle. We compare the two methods from the improvement, feature, and performance. Then we observe the data of each method, and port on Davinci for the Real-time video detection finally. In DaVinci, ARM transfer video data to DSP through share memory for face detection. We cut video window to lots of sub-windows, then we use integral image to compute feature value. If the sub-window can pass all stage, we will search the skin color region whether it is a face window. Finally, we draw the sub-window on image that uses a black square. DSP requires 75ms to process an image and the result output can achieve 13 images per second