Moving Target Classification Based On Similarity Of Feature Model

Moving Target Classification Based on Similarity of Feature ModelWith the rapid development of the information society, people propose a higher demand for the safety of people's life and property:problems about the maintenance of public order gradually attract people's attention. Traditional security measures mainly rely on people, because of human's particularity, they gradually expose the weaknesses. It is hard for people to implement 24 hours of supervision. At the same time, in some special occasions, such as in extreme heat or extreme cold conditions, it is almost impossible for people to supervise at the site. In addition, the cost for employment of security is higher than before. At this point, the video surveillance system displays its superiority. It can basically solve the defects of supervision based on people and provide more secure and reliable security service.From its emergence to now. video surveillance mainly goes through three development stages, analog video surveillance, digital video surveillance and intelligent video surveillance. The surveillance system of every stage is the improvement of the system in last stage, and makes up for the last. With the increasing of human's demand and the development of science and technology, video surveillance moves in the direction towards network, intelligent and high definition. Intelligent video surveillance has some advantages. For example, it can make a real-time analysis and management for the video frames being gathered, and select useful contents captured to store. Then important information is extracted from them and irrelevant information is strained off. So when emergency occurs, it is able to make a response promptly.The major work of intelligent video surveillance could be divided into the following sections:detecting of the moving targets, classification of the moving targets, tracking of the moving targets and behavior understanding. By reading a lot of literature in Chinese and English, making a deep research on the existing motion detection, moving objects classification methods, and related theoretical knowledge, based on a lot of experiments, this article proposes its own method. It is proved by experiments that this method can achieve the desired classification of the moving targets in ordinary outdoor scenes.First, a thorough research on the existing methods of moving object extraction is made. The methods commonly used are:temporal difference, background subtraction, optical flow method and three frame difference. These methods were compared and analyzed through several experiments. In this paper a method which is the combination of three frame difference and background subtraction is proposed. Also, the background updating method is improved. According to the number of white pixels in the sliding window of the images obtained after background subtraction, background model is updated. This method can effectively make up the lack of use of one of the two methods alone, experiments show that this method can achieve an ideal test result and has a wide range of adaptability.Secondly, a preprocessing is made on the images got in the last step that contain moving targets. Through the research and experiments on several common basic operations of morphological processing, for example, erosion, dilation, opening operation and closing operation, according to their respective functions and characteristics, they are used by a certain combination in the work of preprocessing on the foreground objects. This method can make up the incomplete of the moving object's silhouette obtained in motion detection process and the interference caused by background noise. The preprocessing makes a preparation for the work about feature extraction in the target classification.Then, based on the first two steps, the next is to extract the features of the obtained moving targets. By making a research on the existing description methods of image features, such as color features, texture features, shape features and video image features. Video image in relation to the general image has more features, such as motion features. In this paper, the combination of the shape and motion characteristics is used to classify. For the problems in experiment, a new shape feature, namely the number of local peaks, is proposed. The new feature is combined with several existing shape characteristics, through multiple comparisons and experiments, an optimal combination of the characteristics used as the feature vector during the classification can be found. Then the corresponding feature vector models for several known categories, such as single people, vehicle and crowd, are established.Finally, the in-depth analysis and comparisons of the commonly used classification criteria are made and the similarity measure method is used at last. The normalization methods about feature vector and the methods on calculating the distance of similarity are studied. First, Gaussian normalized method is used to dispose the features of the moving objects. According to the size of the contribution in classification of every characteristic component, the weight of each component is set. Then the similar distance between target under test and characteristic vector model is calculated, using the Formula of X2 statistical distance. So target classification is achieved based on the size of the similarity distance.In summary, the main work completed in this paper is on the basis of moving object extraction, to achieve moving target classification based on the similarity of characteristic model. In this paper, a new target motion detection method is presented and a new shape feature is put forward. Through experiments, an optimal combination of the existing features is got. and the similarity measure method is applied to target classification. With the development of society, the technology of moving target classification will be constantly innovated. I hope the work done in this paper can make a little contribution for the development of the whole field.