Vision based Human Activity Recognition using Deep Neural Network Framework

Abstract

Human Activity Recognition (HAR) has become a well-liked subject in study as of its broad application. With the growth of deep learning, novel thoughts have emerged to tackle HAR issues. One example is recognizing human behaviors without exposing a person's identify. Advanced computer vision approaches, on the other hand, are still thought to be potential development directions for constructing a human activity classification approach from a series of video frames. To solve this issue, a deep learning neural network technique using Depthwise Separable Convolution (DSC) with Bidirectional Long Short-Term Memory (DSC-BLSTM) is proposed here. The redeeming features of the proposed network system comprises a DSC convolution that helps to reduce not only the number of learnable parameters but also computational cost in together training and testing method The bidirectional LSTM process can combine the positive and the negative time direction. The proposed method comprises of three phases, which includes Video data preparation, Feature Extraction using Depthwise Separable Convolution Neural Network algorithm and DSC-BLSTM algorithm. The proposed DSC-BLSTM method obtains high accuracy, F1-score when compared to other HAR algorithms like MC-HF-SVM, Baseline LSTM Bidir-LSTM algorithms.