Development of background subtraction algorithm for biometric identification

Abstract

This thesis presents an improved approach for an automatic face detection system. Segmentation of novel or dynamic objects in a scene can be achieved using background subtraction or foreground segmentation. This is a critical early step in most computer vision applications in domains such as surveillance and human-computer interaction. The proposed system consists of three parts. In the first part, the use of background subtraction algorithm to deal with the problem of lighting changes, shadows and repetitive motions. All previous implementations fail to handle properly one or more common phenomena, such as global illumination changes, shadows, inter-reflections, similarity of foreground color to background and non-static backgrounds (e.g. active video displays or trees waving in the wind). The proposed method is a background model that uses per-pixel, time-adaptive and Gaussian mixtures in the combined input space of pixel neighborhood and luminance invariant color. This combination in itself is novel. In the second part, another technique known as morphological erosion and dilation operators are used to remove the noise in the resulting binary image to improve the accuracy. The third part is accomplished by using a new technique to locate the face position in the image and extract ilfor recognition and identification purposes. The algorithm has been tested in several different lighting conditions and environments. The experimental results show that the method possesses much greater robustness to problematic phenomena than the prior state of the art methods, without sacrificing real-time performance, making it well-suited for a wide range of practical applications in video events which requiring detection in real-time. The experimental results in real time applications show the robustness, reliability and efficiency in fhe proposed approach; they can accurately detect and extract human face 98% of the time, with the ability to detect the face of different types of people gender, skin color and head attire. The proposed algorithm can be executed at 30 to 35 FPS for an image size of 320 x 240 pixel, which is much better when compared with any other real time applications.