基於HOG-HOF特徵與動態時間歸整之單次學習手勢辨識

Abstract

手勢影像辨識具有廣泛的應用層面，是電腦視覺中一個重要且活躍的研究領域。本論文基於方向梯度直方圖(HOG)、光流直方圖(HOF)以及動態時間歸整(DTW)，設計出單次學習(one-shot-learning)之手勢辨識系統。本系統分為三部份，包含影像前處理、特徵擷取以及手勢辨識。 首先，本論文針對RGB影片所提出之前處理步驟包含反襯伸展(contrast stretching)與降低影像大小；針對深度影片所提出之前處理步驟包含圖像修復(inpainting)、中值濾波器以及反襯伸展。其次，在特徵擷取部分，所採用之特徵為深度影像之方向梯度直方圖，以及RGB影像之光流直方圖，針對光流直方圖，本論文設計一權重函數，使得Lucas-Kanade光流法的估計結果更為適當。最後，在手勢辨識部分，系統採用動態時間歸整搭配二次卡方距離，同時執行手勢辨識與手勢序列切割。由實驗結果可看出本系統相較於採用相同資料庫之文獻有較好的辨識率。

Vision-based gesture recognition is an important and active field of computer vision research due to its wide applications. This thesis develops a one-shot-learning gesture recognition system which is based on Histogram of Oriented Gradients (HOG), Histogram of Optical Flow (HOF), and Dynamic Time Warping (DTW). The system is composed of three parts, including preprocessing, feature extraction, and gesture recognition. The preprocessing algorithms are proposed for both the RGB videos and the depth videos. For the RGB videos, they are preprocessed by contrast stretching and downsampling, while for the depth videos, they are preprocessed by inpainting, median filter, and contrast stretching. As for the feature extraction, HOG and HOF are respectively extracted from the depth videos and the RGB videos. Besides, a weight function is designed for Lucas-Kanade optical flow model to obtain a proper estimation of optical flow. Finally, DTW with Quadratic-Chi distance is adopted to execute gesture recognition, and temporal segmentation is simultaneously performed. The experiment results show that the proposed system has a better performance when compared to some other approaches applied to the same database ChaLearn Gesture Dataset 2011.