2013 · 23 citations · 28 references
ComputingComputer ScienceComputer VisionHuman BehaviorMachine LearningAction RecognitionDimensionality ReductionMotion AnalysisMotion EstimationPattern RecognitionRepresentation LearningActivity RecognitionBenchmark DatasetsData Redundancy ReductionRandom Walk ModelsSelf-organizing MapsSpatiotemporal Feature ExtractionSpatiotemporal Feature LearningVideo Action Recognition
In this paper, a Dynamic Structure Preserving Map (DSPM) is proposed to effectively recognize human actions in video sequences. Inspired by the latest feature learning methods, we modified and improved the adaptive learning procedure in self-organizing map (SOM) to capture dynamics of best matching neurons through Markov random walk. The DSPM can learn implicit spatial-temporal correlations from sequential action feature sets and preserve the intrinsic topologies characterized by different human motions. A further advantage of DSPM is its ability to learn low-level features in challenging video data. The projection from high dimensional action features to low dimensional latent neural distribution significantly reduces the computational cost and data redundancy in the recognition process. The effectiveness and robustness of the proposed method is verified through extensive experiments on several benchmark datasets.
28
Recognizing human actions: A local SVM approach
2004 · 2.9K citations
Ivan Laptev · International Journal of Computer Vision · 2005 · 2.7K citations · Full text
Behavior Recognition via Sparse Spatio-Temporal Features
Piotr Dollár, Vincent Rabaud, Garrison W. Cottrell et al. · 2006 · 2.5K citations
Computing, Zoology, Computer Vision +14