Effect of contextual information in human action recognition in videos

• Main Author: Koh, Hong Wei
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2016
• Subjects: DRNTU::Engineering
• Online Access: http://hdl.handle.net/10356/67455
• Institution: Nanyang Technological University

With the growing threat of terrorism around the world, many law enforcement agencies have adopted computer vision technique to improve their operation efficiency. Human action recognition has been in the spotlight for the past few years due to its wide variety of capabilities such as gestures-based and surveillance applications. The accuracy of recognizing human actions under different environments have improved tremendously by making use of contextual information to provide relevant clues which further examines human actions. Owing to the element of huge intra-class disparities, it is challenging to recognize the features of every individual. This report focuses on the procedures of capturing and integrating contextual information with the feature covariance matrix to build a framework for analyzing its effect in human action recognition. The first phase of the project is to extract features from the videos using log covariance matrix with Support Vector Machine (SVM) and Extreme Learning Machine (ELM) as the classifiers to discriminate the actions. In the second phase of the project, the motion context descriptor was extracted from every single frame in the videos and integrated with the feature covariance. The combination of both features allows us to achieve an accuracy of over 95%. Additionally, a study into the ability to handle occlusion and body posture variation has been carried out using dense trajectory.