Camera based human computer interaction : Part 1

• Main Author: Yao, Chen
• Other Authors: Chan Kap Luk
• Format: Final Year Project
• Language: English
• Published: 2009
• Subjects: DRNTU::Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: http://hdl.handle.net/10356/17067
• Institution: Nanyang Technological University

This report describes a study of human hand detection based on appearance-based object detection techniques. Inspired by the successful implementations of frontal face detection in cameras, a camera-based hand detection method for human-computer interaction is to be developed. In appearance-based methods the models are learned from a set of training images which should capture the representative variability of hand gestures. The hand detection typically consists of three stages, extracting the features of the training images, training a classifier using the statistical analysis and machine learning of the features, and calculating the correlation of query image with the classifier. Haar-like features are proved to be effective for face detection and they are fast to compute. Hence Haar-like features are calculated using integral image technique. For a 24 x 24 training image, the number of Haar-like feature is over-complete. It takes weeks to train a classifier using so many features. To reduce the feature dimensionality, it is proposed in this project to evaluate the class separability of each feature and only select the best hundred of features to train classifier. Results obtained shows that three-rectangle Haar-like feature has best class separability and lowest test error which is better than other type of Haar-like features. Each feature is a weak classifier and some weak classifiers are boosted to train a strong classifier. To increase the detection accuracy and to reduce the training time, only best class separability features are used for training a classification tree. Results obtained shows that the classifier using a combination of four types of best features performances much better than a classifier using randomly chosen features. The classifier using 20 features, 5 features from each type, could achieve zero validation test error. This perfect result may be due to the small size of the validation set of 200 images. But it stills shows that this method achieves very good test result. This shows that appearance-based approach using Haar-like feature and classification tree using best class separability features is suitable for hand detection.