HUMAN ACTION RECOGNITION USING RGB-DEPTH VIDEOS

Abstract

Being able to detect and recognize human activities is essential for several applications, including personal assistive robotics.Many approaches have been discussed in the past. Normally 2D data has been used in past. But , nowadays due to availabilty of low cost 3D cameras like Kinect, it is easier to perform research on depth data. Mainly skeleton and depth data provides more reliable and accurate system. In this Dissertation, a novel approach to detect the activities performed by a human has been implemented. This involves the extracting the frames from a given depth video and getting the skeleton of human in each frame using kinect camera. Simple skeleton feature are used, which are e cient and fast to classify the activities using multiclass svm.This approach gives a better accuracy in comparison to many approaches developed in the past.