Activity-based methods for person recognition in motion capture sequences

摘要: We present two algorithms for person recognition from motion capture data.The first algorithm is based on a similarity measure between sequences.The second combines dimensionality reduction and Bag of Words approach.Correct rates higher than 98% are achieved.Human activities other walking can be used recognition. In this paper we efficient operating upon data, depicting persons performing various everyday activities. The approach driven the assumption that, if sequences depict certain activity performed by same person, then, consecutive frames (poses) one sequence expected to similar other. proposed method constructs pose correspondence matrix represent poses utilizes an intuitive estimating score sequences, structure matrix. model (BoW), where histograms extracted frequency occurrences characteristic poses. This combined with application Locality Preserving Projections (LPP) in order reduce their dimensionality. Our methods achieved more correct rate, three different datasets.