Abstract

Background subtraction refers to background update and object detection, and it is a commonly used object segmentation technique. In this technique a background model frame is built and updated over time such that it only corresponds to static pixels of the monitored scene. Moving objects are then detected by subtracting each new frame from this background model frame. In this thesis, we propose two real-time effective techniques for video object segmentation: the first is a background subtraction technique that includes background update and object detection stages to extract object binary blobs; the second is an improved contour tracing and a new filling algorithms to extract object features such as area, compactness and irregularity. The proposed background subtraction technique effectively models the static background and detects true moving objects while retaining computational efficiency for the real-time criteria. In the background update stage of the proposed background subtraction technique, the reference background pixels are modeled as multiple color Gaussian distributions (MOGs) with a new selective matching scheme based on the combined approaches of component ordering and winner-takes-all. This matching scheme not only selects the most probable component for the first matching with new pixel data, greatly improving performance, but also simplifies pixel classification and component replacement in case of no match. Further performance improvement to background update stage is achieved by using a new simple yet functional component variance adaptation formula. A periodical weight normalization scheme is used to prevent merging temporary stopped real foreground object into the background model, and the creation of false ghosts in the foreground mask when these objects start to move again. The proposed background update technique implicitly handles both gradual illumination change and temporal clutter problems. The object detection stage uses two schemes that improve object blob quality: a new hysteresis-based component matching to reduce the amount of cracks and added shadows; and temporal motion history to preserve the integrity of moving object boundaries. In this stage, the problem of shadows and ghosts is partially addressed by the proposed hysteresis-based matching scheme, while the problems of persistent sudden illumination changes and camera perturbations are addressed at frame level depending on the percentage of pixels classified as foreground. After background subtraction the detected moving object pixels (initial foreground binary mask) are highly abstract and must be grouped together to form the actual objects. We propose an improved contour tracing and new filling algorithms for grouping object pixels. The proposed improved tracing algorithm can detect and reject dead or inner branches, false non-closed contours, noise related small contours, and then efficiently categorize each contour into inner or outer contours. The new filling algorithm is efficient and never leaks, it uses the extracted contour points and their chain-code information as seed points for horizontal line growing. Experimental results show that the proposed tracing and filling techniques improve computational performance with no tracing or filling errors compared to other reference techniques.