Fast and Robust Background Subtraction
Motivation
Background Subtraction is a standard procedure for higher level applications like People Tracking, Human Motion Estimation, Surveillance, etc.. These applications normally use a static camera observing a scene and are in need of a reliable distinction between fore- and background of the scene on the camera's pixel plane.
The task of this thesis will be the implementation and evaluation of various background modeling algorithms (e.g. Gaussian Mixture Models (GMM) and Local Binary Patterns (LBP)) using GPU processing for real-time applicability.
Abstract
Background Subtraction is a commonly implemented procedure for higher level applications like People Tracking, Human Motion Estimation and Surveillance. As a rst and crucial part in moving object detection it provides a distinction between fore- and background objects in video data that was captured with a stationary camera. Despite the vast pool of methods that were published the last two decades, background subtraction remains a complex task and a gold standard algorithm has not been found yet.
Varying lighting conditions, moving shadows and changing background geometry are only a few difficulties a good algorithm has to deal with. Last but not least many applications require real time capability with a reasonable video resolution. Pixel and color based methods like Gaussian Mixture Models are very popular but reach their limit in difficult lightning environments. Newer approaches overcome those limitations by including spatial and temporal features. The aim of this work was to implement and analyze an algorithm that meets the requirements of an indoor environment.
The algorithm fuses therefore an adaptive Gaussian Mixture Model along with a texture analysis that uses Scale Invariant Local Ternary Patterns, to obtain an expressive feature space that is able to handle shadows and camou aged objects. Sudden illumination changes are detected by an additional post-processing step that is based on luminance histograms. To exploit the power of today's highly parallelized GPUs, the implementation was almost completely ported to the GPU using OpenCL and allows to process full HD resolutions in real time. Finally, the system performance is comprehensively evaluated on relevant real and virtual generated indoor environment datasets.
Requirements
- You should be motivated in AR and should have knowledge of basic image processing algorithms.
- Knowledge of C++ is obligatory.
- Knowledge of OpenCV is auxiliary.
Contacts