| In the field of image processing, there are higher task which is image understanding and lower task such as moving object detection and tracking. As one of the key parts in computer vision field, moving object detection is the basis of video analysis and application, such as content-based retrieval, moving object recognition and tracking, and video surveillance, etc, what's more the basis of image understanding. And it has great significance on application, for instance, the monitoring and classification of moving object (humans, animals or vehicles) in the scene monitored by the camera. Moving object detection provides a classification of the pixels in the video sequence into either foreground (moving objects) or background. At present, background removal, optical flow, temporal difference, Snake model, and watershed algorithm are widely used to accomplish moving object detection. All of these algorithms are readily subject to environment factors such as light condition and shadow interference. Although many scholars have researched these problems and presented some improvement algorithms, these problems are not completely solved.Consequently, a new method of moving object detection combined with image annotation is presented in this paper. It can successfully detect the moving information of original object and new object.The main thought of the moving object detection method represented in this paper are: first of all, a training images set which contains several colorful images after intensity normalization is composed. The color knowledge, namely color codebook, is extracted from the training images set that are trained by the learning algorithm of SOFM neural networks. Second of all, the color quantization of one of the images in the detected scene is finished according to the color knowledge, and the segmentation of the new image is finished by color histogram. The segmentation method is:the color histogram is divided into several areas by the independent peaks gotten from the color histogram, and then the local best threshold is extracted from every sub-area by employing Ostu algorithm. The image is segmented automatically by the local best thresholds. After the image segmentation, the segmented areas are represented by the features, such as Affine Invariant Moments, regional center of gravity, and color matrix, therefore the scene knowledge base is built. Last of all, image segmentation and feature representation are done to the detected image in this scene, and the feature vectors of the objects are matched with the feature vectors of the scene knowledge base by the similarity criterion in this paper in order to detect moving information of the original objects and new objects in the scene. Large amount of simulation experiment data show that moving information of the object can be detected successfully by the moving object detection algorithm combines annotation represented in this paper, and this algorithm has good robustness and practical application. |
