Parallelization of Real Time Video Processing

Abstract

Identifying objects of interest in a video sequence is a fundamental and essential part in many vision systems. A common method to achieve that goal is to perform background subtraction. For automated surveillance systems with multiple cameras, real-time background subtraction is particularly important. In this research, we examine how to exploit GPU parallelism to accelerate the single Gaussian background subtraction algorithm achieving real-time processing of multiple concurrent videos. Experiments performed on a low end GPU showed promising results. The thesis is composed of five chapters together with the table of contents, the list of figures and tables, and the references. Chapter 1 introduces the thesis and summarizes its scope, objectives and contributions. Chapter 2 represents the literature survey of state of art algorithms for background removal in a video frame. Chapter 3 gives an overview on OpenCV library and CUDA C used in our implementation. Chapter 4 presents a detailed implementation of our background subtraction algorithm with different versions of code implemented for both the CPU and GPU; along with a novel of three levels of optimization on the GPU to speed up the processing of multiple videos running concurrently. Results are collected from different experiments and compared with previous researchers work. Chapter 5 gives the work conclusion and Future work.