Human Action Recognition for Intelligent Video Surveillance

DissertationCrime remains a persistent threat in South Africa. This has significant implications for our ability to function as a country. As a result, there is a dire need for crime prevention strategies and measures that seek to reduce the risk of crimes occurring, and their potential harmful effects on individuals and society. Many local businesses, organisations and homes utilise video surveillance as a measure, as it can capture the crime as it is committed, thus identifying the perpetrators, or at least presenting a few suspects. In current video surveillance systems, there is no software that enables security officers to manage the data collected (i.e. automatically describe activities occurring in the video) and make it easily accessible for query and investigation. Access to the data is difficult because of the nature and size of the data. There is a need for efficiently extracting data to automatically detect, track, and recognise objects of interest, including understanding and analysing data through intelligent video surveillance. The aim of the study is to create an intelligent vision system that can identify a range of human actions in surveillance videos. This would offer security officers additional data of activities occurring in the videos, thus enabling them to access specific incidents faster and provide early detections of crimes. To achieve this, a literature study was done in the research area to reveal the prerequisites for such systems, the separate software modules designed and developed and eventually integrated into the intended system. Tests were developed to validate the system and evaluate how all the modules work together. This inevitably confirms the functionality of the fundamental components and the system in its entirety. The results have indicated that each module in the system operates successfully, can effectively extract pose estimation features, generate features for training/ classification and classify the features using a deep neural network. Further results showed that capability of the system can be applied to intelligent surveillance systems and enable security officers’ early detection of abnormal behaviour that can lead to crime