Research And Implementation Of A Non-contacting Remote Guardianship System For The Elderly

With an increasing trend into the aging society of our country,the issue of healthcare and daily life safety for the elderly have aroused much concern and become urgent to be solved.The intelligent video analysis technology has tremendous application value in remote guardianship for the elderly,since it can help to detect,recognize,track people and analysis their behavior,which lays a good foundation for the guardianship of the healthcare and safety for the elderly.Based on the research on some related key technologies,we design and implement a non-contacting remote guardianship system.Firstly,we describe the related research background and the current progress in this field at home and abroad,based on which we introduce the concept of remote guardianship for the elderly,we also elaborate on the detailed framework and key component modules of our system.Secondly,we analyze and compare three universal algorithms for moving object detection.Considering the specific characteristic of the practical scenarios,we choose a background subtraction method based on mixed Gaussian model through extensive experiments.Thirdly,we introduce the basic theory of online object tracking and conduct a thorough research on an object tracking algorithm based compression theory(CT for short).We propose an improved long-time compressive tracker to overcome the drawbacks of the traditional CT algorithm,i.e.,the fixed-size tracking windows and incapability of dealing with occlusion as well as re-identification of the target after a track failure.Through multi-scale appearance matching of the target,tracking windows adaptive to the object size are generated.Besides,the changes of the global feature of consecutive frames in a sliding window are analyzed to see if an occlusion occurs.Finally,to solve the problem of drift,an online detector using HOG features is learned to re-identify the lost object.The experimental results demonstrate that our algorithm performs better than the CT algorithm in terms of both robustness and precision.At last,we carry out research on the motion history image(MHI for short)and Zernike moment.The MHI can describe the history of one motion accurately and the Zernike moment can depict the dynamic local information of one motion.In our experiments,the Zernike moment is extracted from the MHI and used as the description of the human action's feature.The experimental results demonstrate that the proposed algorithm using this method can correctly recognize some simple actions like walking,bending,falling down and so on.We also propose an effective algorithm for falling down detection with the help of the aspect ratio of human contour and effective area ratio.