結合Adaboost與SIFT特徵之手勢辨識系統

Abstract

基於Viola-Jones 技術的現存手勢辨識方法，存在著二個基本的問題。在偵測器訓練過程中，訓練樣本中背景的干擾以及在偵測時物件在照片中平面旋轉對 於偵測結果的影響。像手是一種非固定形物體，訓練的樣本中難免的包含了許多其他的背景物件，如此條件下訓練的偵測器的效能便會降低。現存解決背景物件 影響的方法往往需要花費相當多的人力和計算時間。雖然Viola-Jones 技術下依舊可以使用多個各別不同角度的偵測器去去達到旋轉不變的特性 (rotation-invariance)，但如此便會花費更多的訓練和偵測時間。我們提出了一個基於Adaboost 使用SIFT 特徵的旋轉及尺度不變偵測器。由於SIFT 特徵本身良好的旋 轉不變性、尺度不變性(scale-invariance)和抵抗背景雜訊的特性，我們的偵測器便可以達到旋轉不變性，並可以讓物體以不同大小出現時都能被偵測。分辨不同的手勢的難 度並不比分辨手和背景的難度還要低。因此，我們使用共享特徵(sharing feature)去分辨一張圖像內是否存在一個手勢在裡面。反之，非共享特徵(Non-sharing feature) 可以幫助我們在不同的手勢之間做分類。實驗顯示了一個比起其他方法更佳的訓練速度和分類的準確率。

EXISTING hand detection approaches based on the Viola-Jones’ methods have two fundamental issues, background noise of training images could generate poor performance and rotation-variant. As hands are non-rigid objects, positive training images often contain many other objects which degrade the training performance in Adaboost dramatically. Existing approaches often involve a great deal of manual labeling and a highly computational cost. Although the approaches based on the Viola-Jones’ methods could achieves rotation-invariant in a way of treating the problem as a multi-class classification problem, the process would need more training images and lose training and detection performance. We propose a rotation-invariant hand detector using discrete Adaboost with Lowe’s SIFT keypoint detector, which solves the addressed problems simultaneously. Minimal effort is needed for labeling training data and the performance is maintained. As SIFT keypoints are invariant to translation, scaling and rotation, and are minimally affected by small background noise, the proposed approach achieve rotation-invariant detection straightforwardly. How to create the multi-gesture classification systemis the next step after the single gesture detector is trained. Sequentially executes the single gesture detectors is a general approach to classify multi-gesture. However this approach increases the recognition time as the number of gestures. Classifying different gestures is harder than only classifying a gesture and background. We use sharing features to classify the image is a hand gesture or not. Non-sharing feature can point out the diversity of between gestures and achieve better recognition result. The experiment show a better training and recognition speed and accuracy compared to other existing approaches.