ارائه ي يك توصيفگر جديد براي شناسايي حركات انسان در دنباله ي تصاوير با استفاده از هارمونيك هاي كروي

چكيده انگليسي :

86 New Descriptor for Action Recognition in Video Based on Spherical Harmonics Parvin Razzaghi p razzaghi@ec iut ac ir Date of Submission 2010 10 16 Department of Electrical and Computer Engineering Isfahan University of Technology Isfahan 84156 83111 Iran Degree M Sc Language FarsiSupervisor Maziyar Palhang palhang@cc iut ac irAbstractThe aim of this thesis is to introduce a new descriptor for the action recognition in video Action recognitionis considered as a combination of human action representation and subsequent classification tasks Humanaction representation is divided into two steps First the region of interest is extracted and then an actiondescriptor is used to code this region Human action representation approaches can be either global or local Global representations describe the region of interest as a whole then this region is described with adescriptor The region of interest is constructed by localizing the human silhouette where a backgroundsubtraction or tracking methods can be employed Global representations capture structural information thatmakes the recognition process robust to local deformations In contrast to the global representations localrepresentations describe human action in an image sequence as a collection of spatio temporal features thatcontain interesting information about the action In this thesis we chose global representations due to theirability in modelling structural information specific to an action In addition global representation methodsare more robust to intra class variations for example every person might walk with a different gait Space time volume STV is based on global representation Human motion is completely represented by STV which is constructed over successive frames by stacking human silhouettes in consecutive frames STVcomprehensively contains spatial and temporal information about an action A new invariant action descriptor based upon spherical harmonics is introduced to describe the STV Thegeneralizations of Fourier expansion of periodic functions on the line and polar coordinate representation offunction in the plane to three dimensions lead to the theory of spherical harmonics Spherical harmonic basisfunctions are constructed on a unit sphere with two parameters in the spherical coordinate system Todescribe surfaces regardless of whether they are stellar or not spherical harmonics in its parametric form is x y zused It is shown that the three sets of coefficient clm clm clm can completely define the shape In thisthesis it is discussed how spherical harmonics are invariant with respect to translation scaling and rotation Then we suggest ways to make them fully invariant with respect to these factors Since obtaining an accurate human silhouette is a computer vision problem yet to be perfectly resolved werequire a descriptor robust to noise and outliers It has been shown that spherivcal harmonic coefficients withlower l values describe the overall properties of the shape They correspond to a low pass filter whereas thehigher values add finer and higher frequency detail to the shape So the proposed descriptor is partiallyrobust to noise and outliers We applied the proposed descriptor to the KTH Weizmann Robust and Gesture dataset and compared theperformance of our algorithm to similar methods in the literature The results of our experiments show thatour method performs at least as well as other available methods Also it is shown that our approach is robustto different cycles of the same actions in videos Keywords Computer vision Action Recognition Space time volume Spherical Harmonics

رزاقي، پروانه

ارائه ي يك توصيفگر جديد براي شناسايي حركات انسان در دنباله ي تصاوير با استفاده از هارمونيك هاي كروي

حجم - مكان - زمان , انحراف زماني پويا