A comparative study using improved LSTM /GRU for human action recognition

Abstract

One of the deep learning algorithms for sequence data analysis is a recurrent neural network (RNN). In a conventional neural network, the inputs and the outputs are independent of each other. At the same time, RNN is considered a type of Neural Network where the output from the previous step feeds information to the current phase. It has many applications, including video sentiment classification, speech tagging, and machine translation. Recurrent networks are also distributed parameters across each layer of the network. Several layers are stacked together to increase depth in forwarding and backward information of long short-term memory (LSTM) and Gated Recurrent Unit (GRU). This paper proposes two models for various action recognitions using LSTM and GRU, respectively. The first model was improved by increasing the LSTM layers to four and the number of units in each layer to 128 cells. While in the second model, GRU layers were extended to two layers with 128 cells, and the (update and reset) gates are modified based on the previous and current input. A comparative study was conducted during the experimental tests performed on the UCF101 action dataset regarding the accuracy rate for both models. Test results indicate that the accuracy has a significant improvement compared with other state-of-the-arts action recognitions, which are 95.19% and 92.9% for both improved LSTM and GRU, respectively.