Research On Foreground Object Detection Algorithms In Computer Vision

With the rapid development of information era, computer vision has become more and more popular during the researchers. Computer vision uses computers to analyze and process the surrounding scene via images or videos, trying to simulate the human vision system. Before doing the high-level processing such as recognition, tracking and analysis, computers usually need to detect the foreground targets in images or videos first. In this thesis, the main topics are the related algorithms about the foreground targets in computer vision, specifically the foreground targets detection based on visual saliency detection in images and the foreground targets detection based on background modeling in videos. The visual saliency information of images can describe the different level of stimulation caused by different regions of the images in unconscious situation, so it can be used as the basis of foreground detection in images. Background modeling is the most widely-used method for segmentation of foreground in video image sequences. The accuracy of background model will greatly affect the accuracy of the following processing.So, according to these two basic issues above, saliency detection algorithm based on wavelet transform and region fusion and background modeling using local sample density outlier detection are proposed in the thesis.◆ Saliency detection based on wavelet transform and region fusion, first we use wavelet decomposition and reconstruction to gain the feature maps in different channels and different scales, calculating for the local saliency maps; than we use Mean-Shift to segment and cluster the input images, fusing the similar regions, and gain the global saliency maps by calculating the color distance of different regions; using global salient values as the region weights and modulating to the corresponding local salient values to gain the final saliency maps. We compare the saliency maps gained by different saliency detection models and calculate the precision, recall and F-Measure, to indicate that our saliency detection model has◆ more advantages and can get better results.◆ Background modeling using local sample density outlier detection, first we calculate the local background factor LBF of every sample. We compare the LBF of newly observed pixel to the LBF of its neighborhood to decide whether it belongs to the background or not. The background modeling method can successfully handle different situations, especially for the scenes containing dynamic backgrounds, we can get better results comparing to other background modeling algorithms.