Evaluation of MoG Video Segmentation on GPU-based HPC System

Abstract

Automated and intelligent video surveillance systems play an important role in the modern world. Since the number of various video streams that must be analyzed concurrently grows, such systems can assist humans in performing tiresome tasks. In order to be effective, video surveillance systems have to meet several requirements: they must be accurate and able to process the received video stream in real-time. A robust system should not depend on lighting conditions, illumination changes and other sources of scene variation. A common component of surveillance systems is a module that performs background estimation and foreground segmentation. The MoG (Mixture of Gaussians) algorithm is a widely used statistical technique of video segmentation. The estimation process is time-consuming, especially for complex mixture models containing many components. The work presented here focuses on the performance evaluation of MoG algorithm aiming to assess feasibility of OpenCL-based processing of high resolution video on GPU accelerated platforms.