An Optimized Deep LSTM Model for Human Action Recognition

The behaviour of human is termed to be an imperative aspect in social communiqué .The detection of human activities represents a type of clues that provide assessment of human behaviour.The recognition of human activities is complex because of the large alterations of human activities in day-to-day life.Also, the accurate action recognition is a complicated procedure due to cluttered backgrounds and changes in viewpoint variations.This paper designs a technique to identify the actions of humans using optimized Deep Long Short Term Memory (Deep LSTM).The aim is to devise an optimization driven deep model for determining the actions of human considering a set of videos.The extraction of video frame is performed.Then, the features, like spider local image feature, shape local binary texture (SLBT), local Texton XOR pattern, Local Gabor Binary Pattern (LGBP), Shape Index histogram, Local Gabor XOR patterns (LGXP) and statistical features are mined.After that, the detection of human action is done using Deep LSTM wherein training is implemented with proposed improved invasive weed based Poor rich (IIWBPR) algorithm.The proposed IIWBPR-based Deep LSTM outperformed and provided supreme accuracy of 92.3%, sensitivity of 92% specificity of 92.6% and F1 Score of 91.9%.The accuracy of the IIWBPR-based Deep LSTM is 17.77%, 15.06%, 8.02%, and 7.80% improved than the existing comparative methods.