The Research Of Moving Target Classification Algorithm Within Specific Area Based On Multi-feature Fusion

With the development of computer technology, the intelligence of video surveillance system is constantly accelerating. The techniques involved in intelligent video surveillance play a vital role in its intelligence. Among them, moving target classification plays an important role of connecting link, on one hand, it accepts the output of target detection, on the other hand, it provides the necessary input for the target behavior understanding and analysis. With the growing application of intelligent video surveillance system, the target classification based on semantic draws more attentions of researchers. Its research and development largely determine the development of automatic understanding of video. Therefore, the study of moving target classification has the profound significance.In this thesis, the existing moving target classification algorithms and the involved theoretical knowledge have been deeply studied. Since the concept of target classification involves some uncertainty, such as:scene、the target categories、 characteristics and so on. We will do some general definition and description. In this thesis, the scene is road monitoring under a fixed background, the categories of targets conclude pedestrian、bicycle/car battery and car. And feature items combine static characteristics with the proposed dynamic characteristic, and a concept is introduced that classification within a manually framed quadrilateral area from the monitor picture. Every stage work of classification algorithm can be described as follows:(1)Several classical moving target detection methods have been studied, and their algorithm and implementation steps have been explained. Among them, background subtraction is highlighted. For three kinds of common background modeling algorithms(SAM、SGM、GMM), experiments and comparative analysis are conducted. Finally, we discuss the advantages and disadvantages of each detection algorithm and its application scope.(2) We use several common image processing methods to improve the shape、 contour information of target, and use it to extract target. The methods include morphological operations、connected domain identifier and the merging of clumps. Accordingly, the extracted features have more complete information about their shapes. This thesis uses a multi-feature fusion method which combines the static characteristics with proposed dynamic characteristic (The lower third aspect ratio variation) to identify the target, it effectively improves the accuracy of target classification. The dynamic characteristic based on rigid and non-rigid principle of target motion reflects the periodic of movement(3)The last step of target classification is to construct a classifier to derive classification results. A good classification algorithm can accurately identify the category of targets according to limited samples and features. This thesis uses SVM to construct the classifier which is suitable for small sample classification. It includes two parts:training and testing. First, constructing a reliable classifier by training samples, then testing the validity of the classifier by test samples, and finally obtaining test results.(4) After the implementation of traditional moving target classification algorithm, considering the application requirements of specific scene, the thesis proposes a classification method based on the specific area. It is useful to extract the target within the interesting area from the wide monitoring scenario. This method needs us to manually calibrate the region of interest, and set it as the scope of detection and classification. Then process the pixels in the framed area by using detection, feature extraction and classification algorithms. This method is mainly used in some special scenarios, such as:the zebra crossing, crossroads and entrances, etc.