Vision-based Human Pose Estimation In Smart Classroom

It is always a dream of AI researchers to enable the computer to understand human behaviors. Human behavior analysis is a technology that integrates computer vision and pattern recognition, which will have broad application in our daily life in the recent future. It will play an important role in military, health care, security, entertainment and Human-Computer Interaction areas. This work proposes future applications of human behavior analysis in the"Standard Nature Classroom", a creative framework of smart space, of the E-Learning lab of Shanghai Jiao Tong University. SNC makes use of a great number of AI and computer vision methods to use the state-of-art computer technologies to support a human centric, intelligent education environment for teachers and students over the network and physical classrooms. One of the key problems in human behavior analysis is the"Human pose estimation"problem. In the paper, we aim at researching and developing a set of algorithm and framework that can output a good quality human pose estimation providing the images of the students and teachers in SNC.The main contribution of this article is proposing a set of human pose estimation algorithms that are based on stereo vision, combining depth information and classical computer vision algorithms. We proposed an improved SAD algorithm to calculating disparity between two images which uses median filter and surface constraints to remove artifacts. Based on the disparity map, we proposed a method that combines depth information and active contour algorithm to extract human body contour from images. Such method is superior to mixture of Gaussian model when the background of the image is complex and changing. Moreover, active contour could output smooth contour which fits better to the human body. Finally, the article experiments two types of pose estimation algorithms that are applied to different scenarios. One of them is a discrete pose estimation algorithm that is based on contour matching. We use Hu invariants to describe characters of a contour and use SVM to train classifier on the training dataset. We achieved 80.7% classification precision on the testing set. The other one is a human pose estimation algorithm based on Pictorial Structure model, which is used to estimate non-occluding human body poses.The paper also describes the design and construction of a prototype system that realize the algorithms to validate the effectiveness of the proposed algorithms and to build a base framework for future researches.