3D Human posture estimation and tracking by using vectorization and time series prediction

Abstract

The tracking and pose estimation of people in video made challenging problem due to the variabilityof human appearance. In this thesis, we research on articulated 3D human motion and pose estimation.However the human posture recognition has many dicult problems to be solved, because the humanbody model is deformable object, besides we have to process the complexity of natural scenes and thehigh dimensionality of articulated body models. Also the several problems such as texture of clothes,lighting change, illumination, various colors made the problem harder.Recent approaches to estimating and tracking human posture exploit articulated human body modelsin which the body is viewed as a kinematic tree. Since the particle filter has been proven as very successfulalgorithm for non-linear and non-gaussian estimation problem, the particle filtering was alternativemethod for human body tracking. Our goal is to develop a modified particle filter which is shown tobe effective at searching the high-dimensional configuration spaces (40 dimensions) by using time seriesprediction. Our results was suggested in constrained laboratory environments and showed our methodsperform quite well than others. Multiple cameras and background subtraction, however, are required toachieve reliable tracking performance. In this thesis, we proposed the admix vector and time series vector(TSV) particle filter. The proposed likelihood which is admix vector can reduce the computation timeand solve the asymmetric problem at the same time. To solve the problem of the robust tracking of fastmoving object, the time series prediction was employed to predict states following regression informationof RBF neural network.