Learning sparse representation via spatio-temporal smoothing for human activity recognition
Date of Issue2014
School of Electrical and Electronic Engineering
Recent years have seen popularities of sparse coding in many research fields. One of these fields is computer vision, where sparse coding has been applied in the process of feature quantization and selection. Although the general sparse coding method reduces the complexity of coding process (hence saves memory space), and makes the reconstruction of the feature from the sparse codes easy, the data that feed into the coding process are not in the optimal state and can cause errors in the subsequent processes. In this dissertation, we propose a new graph-based sparse coding model that optimizes the human activity feature to improve the accuracy of human activity recognition. We demonstrate how exactly the model can optimize the data by using correlation computation. We achieve encouraging performance gains after using this new model. We also compare and discuss three methods for sparsity estimation of feature coefficients. In the end, we find the optimal parameter settings for features, dictionary size, etc. for human activity recognition based on KTH and HMDB51 video datasets.
DRNTU::Engineering::Electrical and electronic engineering::Electronic systems::Signal processing