Compressive Sensing of time series for human action recognition

Compressive Sensing (CS) is an emerging signal processing technique where a sparse signal is reconstructed from a small set of random projections. In the recent literature, CS techniques have demonstrated promising results for signal compression and reconstruction [9, 8, 1]. However, their potential as dimensionality reduction techniques for time series has not been significantly explored to date. To this aim, this work investigates the suitability of compressive-sensed time series in an application of human action recognition. In the paper, results from several experiments are presented: (1) in a first set of experiments, the time series are transformed into the CS domain and fed into a hidden Markov model (HMM) for action recognition; (2) in a second set of experiments, the time series are explicitly reconstructed after CS compression and then used for recognition; (3) in the third set of experiments, the time series are compressed by a hybrid CS-Haar basis prior to input into HMM; (4) in the fourth set, the time series are reconstructed from the hybrid CS-Haar basis and used for recognition. We further compare these approaches with alternative techniques such as sub-sampling and filtering. Results from our experiments show unequivocally that the application of CS does not degrade the recognition accuracy; rather, it often increases it. This proves that CS can provide a desirable form of dimensionality reduction in pattern recognition over time series.