Human Action Recognition Based on Deep Neural Networks

Abstract

To better handle human action data of text type, an accurate and fast human action recognition scheme is proposed in this paper. The scheme fine-tunes two classical deep neural network (DNN) models (ResNet-50, EfficientNet-B3) to recognize different actions by processing text-type data of 19 human actions. The method's feasibility is demonstrated by several sets of experimental data showing that the highest 98.98% Top-l ACC is achieved on the noise-free validation set. The dataset containing Gaussian noise achieved 97.81% and 97.50% Top-l ACC on the validation and test sets, respectively, proving the excellent generalization performance of the method. Applying dropout techniques with different ratios to the fully connected layer of the model, the method is proved to have an excellent ability to prevent overfitting through diverse experiments introduced on the Gaussian noise validation set and test set.