HUMAN ACTIVITY RECOGNITION IN REAL TIME USING DEEP LEARNING

Abstract: In today's world, Human Activity Recognition [HAR] plays a critical role in 'human- to-human' interaction. HAR displays and provides the identification of a human as well as the action done by that human, which is tough to recognize. Due to the high processing time, deep learning techniques such as CNN and LSTM cannot be used, instead we will apply transfer learning to recognize human activities. For many computer vision-based applications, such as video surveillance, criminal investigations, and sports applications, human action recognition is one of the difficult issues. Using the similarities between each pair of frames, each extracted sub-unit is further separated into frames that represent action. We will detect the action by comparing the generated HOG to the existing HOGs in the training phase, which represents all the HOGs of many actions using a dataset, utilising the Histogram of the Oriented Gradient (HOG) of the Temporal Difference Map (TDMap) of the frames.