Yoga Detection using Human Pose Estimation

Abstract

There is an increasing requirement for real-time human pose estimation from monocular RGB images in applications such as human-computer interaction, video surveillance, people tracking, activity recognition, and motion capture. Human pose detection plays an important role in human activity recognition. HPM is fast growing and lately steps ahead with the release of the Kinect system. CNNs with spatiotemporal 3D kernels (3D CNNs) can directly extract spatiotemporal features from videos for action recognition. The aim of this thesis is to show the ways to detect human body parts using 3D CNNs based on ResNet toward a better action representation. This presents a method for real-time multi-person human pose estimation from video by utilizing convolutions neural networks. In this thesis will be the project flow and implementation of the detection of the Yoga and classification steps.