Human Action Recognition by using Dense Trajectories

Abstract

Human activity recognition in video is the major problem in the computer vision. The human detection and tracking is significant challenges in the field of computer vision and pattern analysis which is used for video surveillances, gait pathologies recognition, robotics, human computer interaction and sport. When the video dense trajectories can be used to detect the fore ground motion by using the optical dense flow method. The dense representation is capturing the local information in the each video. The human motion can be tracking by using the spares Kanade Lucas Tomasi based tracker method.

The motion and structure descriptors are used to describe the human motions based on the motion boundary descriptors such as histogram of gradient and histogram of optical flow. These descriptors are detecting the human by computing derivates in the horizontal and vertical components of the optical flow. We evaluate the pseudo likelihood estimation of each trajectory that is shape factor and scale factor values in the each video dataset. The human activity recognition is classifying the human action by using K Nearest Neighbor classifier and binary Support Vector Machine classifier. The MATLAB tool is used to implement the human action recognition. Various types of human actions are tested and the performance been analysed. In this project, experimentally three datasets are used such as Human 3.6Million dataset, KTH dataset and Berkeley MHAD dataset.