Matching mixtures of curves for human action recognition

摘要: A learning-based framework for action representation and recognition relying on the description of an by time series optical flow motion features is presented. In learning step, curves representing each are clustered using Gaussian mixture modeling (GMM). a probe sequence also GMM, then projected onto training space matched to learned non-metric similarity function based longest common subsequence, which robust noise provides intuitive notion between curves. Alignment mean performed canonical warping. Finally, categorized with maximum nearest neighbor classification scheme. We present variant method where length reduced dimensionality reduction in both test phases, order smooth out outliers, these type sequences. Experimental results KTH, UCF Sports YouTube databases demonstrate effectiveness proposed method.