Deep Learning-Based Real-Time Multiple-Person Action Recognition System

Abstract

Recognition of human activity has attracted growing interest from researchers, mainly because of its potential in areas such as surveillance videos, robotics, interaction with human computers, user interface design, and multimedia. Since it includes details about a person's identity, personality, and psychological state, it is difficult to remove. The human ability to recognise another person's behavior is a primary subject of study in the scientific fields of computer vision and machine learning. People's actions are determined by their behaviors, which makes assessing the problem of determining the underlying action extremely difficult. The proposed system would identify and track a large number of people arriving on the scene before recognizing their behavior. Due to the improved resolution of the video frames, when persons in the scene become too far away from the camera, we build a zoom-in feature to produce more suitable action identification results.The work proposed is based on the behavior recognition model of Deep Belief Networks and Deep Boltzmann Machines.