Exploration of Deep Learning Models for Video Based Multiple Human Activity Recognition

Abstract

Human Activity Recognition (HAR) with Deep Learning is a challenging and a highly demanding classification task. Complexity of the activity detection and the number of subjects are the main issues. Data mining approaches improved decision-making performance. This work presents one such model for Human activity recognition for multiple subjects carrying out multiple activities. Involving real time datasets, the work developed a rapid algorithm for minimizing the problems of neural networks classifier. An optimal feature extraction happens and develops a multi-modal classification technique and predicts solutions with better accuracy when compared to other traditional methods. This paper discussing on HAR prediction in four phases namely (i) Depthwise Separable Convolution with BiLSTM (DSC-BLSTM); (ii) Enhanced Bidirectional Grated Recurrent Unit with Long Short Term Memory (BGRU-LSTM); (iii) Enhanced TimeSformer Model with Multi-Layer Perceptron Neural Networks classification and (iv) Filtering Single Activity Recognition are described.In comparison to previous efforts like the DSC-BLSTM and BGRU-LSTM classifications, the experimental result of the ETMLP classification attained 98.90%, which was more efficient. The end outcome revealed that the new model performed better in terms of accuracy than the other models.