A Novel Hybridization Model for Human Activity Recognition using Stacked Parallel LSTMs with 2D-CNN for Feature Extraction
Abdul Manaf F, Sukhwinder Singh · 2021
Recognizing human activity is a difficult task in certain applications, including video retrieval, human-computer interactions, autonomous driving, video surveillance, medical field, educational purpose, and abnormal activity recognition etc. However, extracting reliable and timely information from video recordings of human actions and behaviours is the most challenging issue in a ubiquitous computing environment. The purpose of this work is to present a novel hybridization deep model for comprehending and interpreting videos in HAR systems. The proposed hybridization model is built using 2D-Concolutional neural networks (2D-CNN) and stacked parallel bidirectional Long short-term memory (SPBD-LSTM). Its capable of learning prolonged sequence and processing longer videos that are sequentially synchronized by assessing attributes over a specified time interval. Experiments are carried out with two benchmark datasets UCF101 and SPHAR (Surveillance Perspective Human Action Recognition) to validate the contributions of proposed work. Finally, our hybridization model exhibits outstanding performance for analysing complicated human activities.