Human action recognition using spatial-temporal analysis.
MetadataShow full item record
In the past few decades’ human action recognition (HAR) from video has gained a lot of attention in the computer vision domain. The analysis of human activities in videos span a variety of applications including security and surveillance, entertainment, and the monitoring of the elderly. The task of recognizing human actions in any scenario is a difficult and complex one which is characterized by challenges such as self-occlusion, noisy backgrounds and variations in illumination. However, literature provides various techniques and approaches for action recognition which deal with these challenges. This dissertation focuses on a holistic approach to the human action recognition problem with specific emphasis on spatial-temporal analysis. Spatial-temporal analysis is achieved by using the Motion History Image (MHI) approach to solve the human action recognition problem. Three variants of MHI are investigated, these are: Original MHI, Modified MHI and Timed MHI. An MHI is a single image describing a silhouettes motion over a period of time. Brighter pixels in the resultant MHI show the most recent movement/motion. One of the key problems of MHI is that it is not easy to know the conditions needed to obtain an MHI silhouette that will result in a high recognition rate for action recognition. These conditions are often neglected and thus pose a problem for human action recognition systems as they could affect their overall performance. Two methods are proposed to solve the human action recognition problem and to show the conditions needed to obtain high recognition rates using the MHI approach. The first uses the concept of MHI with the Bag of Visual Words (BOVW) approach to recognize human actions. The second approach combines MHI with Local Binary Patterns (LBP). The Weizmann and KTH datasets are then used to validate the proposed methods. Results from experiments show promising recognition rates when compared to some existing methods. The BOVW approach used in combination with the three variants of MHI achieved the highest recognition rates compared to the LBP method. The original MHI method resulted in the highest recognition rate of 87% on the Weizmann dataset and an 81.6% recognition rate is achieved on the KTH dataset using the Modified MHI approach.