Moving Object Detection In Complex Scenes

Moving object detection in image sequences is an important research domain of computer vision, and it's a significant step for further processing to detect and extract moving objects from time-varying scenes rapidly and accurately. In this paper, our goal is to detect and extract moving objects accurately from real image sequences acquired using a fixed camera and to remove shadows of moving objects.This paper firstly analyzes and summarizes main approaches for motion detection, and then uses lump matching, temporal difference and background subtraction in experiments, and discuses what conditions each of them is suitable for.This paper focuses on background subtraction approach, and proposes a new background modeling approach based on temporal division and merging. This approach works as follow: for each pixel of the image, through a training process, we put points of image sequences to different groups with the points of each group has similar values, and we take the feature value of the group that has the largest number of points as the background feature value of the pixel. Our experiments show that this approach works very well with scenes without complex motions in the background. To make the system flexible enough to handle all kinds of variations to the observed scene, this paper introduces and realizes GMM for background modeling. Results show that GMM works efficiently in most complex outdoor scenes and meanwhile realizes on-line detecting.This paper uses an improved deterministic non-model based with color exploitation which distinguishes moving objects and their shadows in an accurate way, and improves the performance of moving object detection algorithm.Our experiments show that background modeling based on temporal division and merging results in a reliable and robust motion detection system that deal with scenes without complex changes in the background, and causes low computational intensity. But concerning with backgrounds that have unpredictable events GMM is more competent since it can deal with visual noisy efficiently.