基於JointBoost之強健且即時多類別偵測器

Abstract

本研究的目的為建置一高效的多類別偵測器，用以同時偵測並分辨場景中的不同物件。 常見的傳統方法是，單獨訓練多個偵測單一物件的偵測器，對同一畫面做運算。 再者，傳統方法的運算複雜度會隨著所需偵測物件種類的增加而大幅上升。 在本研究中，我們利用AdaBoost演算法選取有效特徵的能力，來組成特徵候選群，以提升準確度。 後續使用JointBoost演算法訓練出可以共享特徵的多分類弱分類器群，並以串接方式將這些弱分類器群組成一個多類別偵測器，以降低多分類器之運算量。 更進一步使用可傳遞式的疊接方法來組成我們的多類別偵測器，因為可將前級結果累進至下一級，不會因為前級的判斷而遺失，亦可提升偵測器準確度。

This study aims to build an efficient multi-class detector that can detect and recognize different kinds of objects in the scene. Traditional approaches usually apply a series of different detectors to the scene, and each detector is trained individually. Moreover, when detected objects increase, the computational load of traditional approaches increase rapidly. In this study, we take advantage of the AdaBoost algorithm to select effective features as a feature pool which improves detection accuracy. JointBoost algorithm is applied to train a multi-class weak classifier set with sharing feature. Muti-class classifiers are then built with a cascade schema to reduce computational load. Furthermore, we proposed a distributed accumulation method to increase classification accuracy of multi-class cascade classifiers. Experimental results show that the proposed method has high object detection rate as AdaBoost and high classification accuracy close to JointBoost with lower computational load due to the proposed cascade schema.