A FRAMEWORK FOR HUMAN ACTIVITY RECOGNITION USING DEPTH AND SKELETON EVIDENCE

Abstract

In today’s world of advancement, computer vision is gaining popularity in different spheres of life. Many areas like imaging, prediction of events, identification, encryption of different languages, etc are covered under it. Activity recognition is one of its area of application. Human activity recognition is helpful in prognosticating the actions of a person or a club of people of different views. The particular topic is inspired from realworld submissions, for example, visual observation, video understanding etc. The main building blocks of activity recognition consist of pre-processing, features extraction and representation, and classification. Considering the various applications of activity recognition, this thesis investigates human activity recognition approach based on human skeleton and its associated depth images. The joints extracted from skeleton representation used in feature extraction and mapped on associated depth images. To achieve higher recognition accuracy of human activities, a three- step methodology is devised: First step is obtaining the skeleton representation to estimate joint location, second step is extraction and representation of features, which is done using two major approaches: first is the histogram based and the second is the vector quantization using linear discriminate analysis and finally is the classification of human activities performed by the neural network. This research is mainly focused on proposing neural network for extraction and representation of features in activity recognition methodology based on skeleton and depth information. The projected methodology is based on two public databases. It showed great accuracy on public dataset.