Human Activity Recognition Using OpenPose and Neural Network

Abstract

Human Activity recognition are the most implemented area of Machine learning and Neural Networks. Previous studies in this area show how Human activity recognition system perform on images of where poses of a person need to be detected, Mostly the studies in this specific research domain work on Single-person pose estimation, as Multi-Pose estimation is a daunting challenge to implement correctly. I proposed a Robust Human Pose and Activity Recognition System for 2D Image/Video based on Realtime input using OpenPose. The open source library is fully implemented and ran to track/Detect/Highlight Human poses effectively. All the pre-requisites, libraries, and software’s which were needed to run OpenSource were merged together to create a single unit creating a standalone Pose recognition system. I will be proposing Low-level image pre-processing techniques which will enable images having low-light conditions to perform exceptional better. This will be optimized according to the Input requirement of OpenPose system. Furthermore, for comparison TensorFlow activity recognition has also been installed to showcase and compare the robustness and fluidity in recognition by OpenPose library. In this thesis i will demonstrate how heat map of Hand, Face and body index are categorized by OpenPose. Different outputs and inputs are compared to show the superior recognition capability of OpenPose Human Activity Recognition. Additionally, Future work has been proposed in this thesis to showcase real-life implementations that are possible through automation of proposed methodology.