The Research And Implementation Of Human Action Recognition Based On Improved ISA Deep Network

The research of human action recognition becomes the hot-spot issue of computer vision. In view of the complex environments, and the confusing similarities of actions, great improvements still can be achieved in recognition accuracy. Recently, with Deep Neural Networks, the accuracy of action recognition has been considerably improved.The topic of this thesis is the research and implementation of human action recognition based on Improved ISA Deep Network. The main works are as follows:1. Analyze the defects of ISA Neural Network and propose the Improved ISA Deep Network based on Subdivision-Fusion Model(SFM), including the training stage based on Subdivision and the recognition stage based on Fusion. Two methods to determine the numbers of subclasses in the training stage are provided, by considering the overlapping feature distribution and the class imbalance problem.2. With Improved ISA Deep Network, the implementation of human action recognition research system based on Improved ISA Deep Network is done.3. Experiments are thoroughly conducted on three popular international standard human action datasets: Hollywood2, KTH and YouTube, and analyses are elaborated.The main achievements and contributions of this thesis are as follows:1. The Improved ISA Deep Network based on Subdivision-Fusion Model(SFM) is novelly proposed. Utilizing the ideas of “SSC clustering and subdivision”, “the remapping from feature space to sample space” and “the second training stage”, a unique mapping transformation which is different from most neural networks is formed.2. The Improved ISA Deep Network can avoid “overfitting”. Training by focusing on more concentrated samples, the network is made deeper indirectly, leading to more abstract and invariant features, and result in improved action recognition accuracy.3. The experimental results indicate that, compared with ISA Neural Network, Improved ISA Deep Network obtains obvious accuracy improvements on Hollywood2, KTH and YouTube datasets, outperforming or reaching the state-of-the-art accuracies.