Smart Yoga

Abstract

Yoga Recognition is a task of Action Recognition. Action recognition is one of the common Computer vision fields which is quite prevalent and useful in various activities like video surveillance, VR games, sports, medical tests and fitness Our Project is about Yoga Detection which can be achieved by various Action recognition algorithms. Action recognition is done by many ways, using various convolutional networks e.g. Imagenet, 2 stream 3D CNN, Activitynet and 3D CNN Resnet based .From all the various architectures we choose 3D CNN Resnet based which extracts Spatio temporal features from the input videos. We started our project from pose estimation, detected key points of various body parts, detected Yoga from the key points now we have moved to Action Recognition using 3D CNN which will provide a more accurate output and dependency on camera angle, side facing camera, overlapping of parts can be decrease substantially. We have custom trained our model with our own dataset which contained 100 videos (70 training 30 validation). Finally we have tested the model for various test videos and noted accuracy. Final Output is video containing name of the Asana along with correction if needed.