The Experimental Comparison of Features for Hand Detection

摘要

Hand detection is critical in gesture recognition for conveying information or control commands between persons and computers. The accuracy of hand detection from images plays an important role in these applications. Extraction of effective features is the main factor in this task. The features should be discriminative, robust to different variations and easy to compute. This paper presents the experimental comparison of features commonly used in object detection, such as Haar-like features, a histogram of oriented gradient (HOG), and local binary pattern (LBP), using hand detection as the test platform. The adaptive boost (AdaBoost) cascade classification method is employed to combine 'weak learners' to a strong classifier. The classifier was trained using 300 positive images, which are images containing the hand (region of interest (ROI)) and 10000 negative images, which are images that do not contain a hand on them. Different parameter combinations of the classifier are considered for comparative experiments. The performance of the classifier using Haar, HOG and LBP features were evaluated with 320 static test images. The results show that parameter combinations have significant effects on the hand detection accuracy, which also differ when different features are used.