Development of Adaptive Sampling Based strategy for Human Activity Predictions Using Sequential Networks
Gaurav Kumar Yadav, G. C. Nandi · 2020
The prediction of human activities dynamically is a formidable issue at the edge of computer vision and graphics. This paper presents an innovative idea of developing an adaptive sampling-based strategy to recognize and predict human activities dynamically. More specifically, we propose a new adaptive cost function to train a Gated Recurrence Unit (GRU) using Human 3.6M sequential data [1]. Subsequently, after getting trained on our proposed cost function, the deep neural network can predict human motions more accurately with a significant reduction of mean angel error than the state of the art models in this area. The adaptive sampling-based cost function is a bell-shaped function, and it takes input from its samples during training to make it capable of learning from its own mistakes during training. We develop and configure this function with several variations of the bandwidth. We analyze and compare our results with several models, such as Long Sh ort Term Memory (LSTM) with three layers, structural Recurrent Neural Network(RNN), Zero velocity, and Residual supervised models. In our model, the convergence of training loss and validation loss is much faster; the reasons for which and other details have been elaborated in this paper.