A Human-computer Interaction System Using Single Camera

Human-Computer Interaction (HCI) is a research on how to communicate between human and computers, and how to provide information management, service and trans-action for users conveniently. Nowadays, computer-centered HCI has been changed into user-centered. Human action is an natural, direct and easy way to operate computers. No medium is needed if we use human body as input device when communication. User can control machines simply with proper predefined actions. So input with human actions is important in HCI. Because of the development of Computer Vision and Digital Image Processing, vision-based action recognition has become an active research in HCI system.This paper focuses on vision-based HCI with human action recognition, using single camera. We discuss techniques in Computer Vision, Digital Image Processing and Pattern Recognition. According to requirements of applications, we improve and integrate some methods.Firstly, this paper gives a full summarization of main steps and related domain on human action recognition system. Current common methods in action recognition are analyzed and discussed. According to the framework and process of common recognition system, we give a survey and overview on related methods clearly. Newcomers of this research field can get a general outline of action recognition from this paper.Then, we expatiate on related work from following aspects:Background Subtraction is the essential step of vision-based action recognition sys-tem. We summarize the basic methods of Background Subtraction, and introduce Back-ground Difference and Mixture of Gaussians in detail. Furthermore, we propose a Two-layered Background Subtraction for removing shadows.Parameterize of human body is the way to transfer complicated contour image to parameters which are easy to recognize. We summarize the main methods of parameter-ize, and analyze suitable cases for each method. We talk about Star Skeleton in detail according to the requirement of Human-Computer Interaction system.Action Recognition is the key module of this system. Common methods are divided into two classes:example-based and learning-based. We analyze both classes, and choose Star Distance and Support Vector Machine as instances to talk about respectively. After the system estimates posture from every frame, we use a Posture Buffer to match posture sequence with action behavior.The innovation point of this paper is improving some basic algorithms, and increasing the performance by integrating new researches. We also implement a practical interaction system. The detailed innovation as following:We propose a Two-layered Background Subtraction method, which is based on both chromaticity and gradient in image. This subtraction helps us remove shadows from human contour. This makes our system more robust to the environment, and improves the usability and stability.We use a novel method integrating Star Skeleton and Support Vector Machine in recognition step. Interaction system requires high performance. We use Star Skeleton to abstract simple parameters from human action. These parameters are easy to recognize and keep main features. Use this model as feature space of SVM can achieve a efficient and promising recognition result.With the improved algorithm, we implement a whole system of human action recog-nition and achieve real-time interaction. A large amount of test cases and experimental results show the promising performance and effectiveness. In addition, this system frame-work can be general used and easy to extend. It has potential to be improved.