Research On Multiple Moving Objects Classification From Video

With the development of computer technology, digital image processing technology, intelligent video surveillance has become an important field of computer vision research. Moving objects classification is an important part of intelligent video surveillance, it is the basis for behavior analysis and understanding. In this paper, after moving objects from video are accurately extracted, a few valid objects features are extracted, the moving objects are accurately classified as persons, cars and bicycles.For detection of moving objects, moving objects detection algorithm based on frame difference and background substraction is adopted, the algorithm contributes to a more complete extraction of moving objects, lays the foundation of accurate classification; Surendra background updating algorithm settles the interference factors appearing in the experiment of the moving objects extraction: the movement of stationary objects in the background and the longly stay in the background of moving objects.For extraction of the moving objects, foreground images are binarized by dynamic threshold segmentation method based on the largest variance between-class in this paper, a more complete moving objects pixels are obtained; image noises are removed by morphological filtering, and then eight-neighborhood connected domain identification algorithm and fragments combination are used to get the complete moving objects.For the moving target classification, the existing common objectives features are refined and a classification method based on dynamic features and static features including aspect ratio,void volume,the change of void volume are proposed in the paper; DAG-SVM multi-class classifier is selected as our classifier; considering the guesture diversities of persons, cars, bicycles, training samples and testing samples including kinds of guestures are selected in this paper, the training samples are inputed to the multi-class classifier to train and the testing samples are inputed to the multi-class classifier training by the training samples already to get the final classification results. The features such as the aspect ratio, the change of void volume have a certain robustness to empty, which can not be solved in some case, accurate objects classification is also achieved.Experiments indicate that the method chosen in this paper can achieve accurate classification of moving objects and obtain good classification accuracy.