Object Detection And Classification In Intelligent Surveillance System

In recent years, people pay more and more attention to our living environment, pay attention to coming down with the prevention of the calamity and foundation of the urgent counter-measure , hope to improve production and the intelligent level of management lives urgently , the intelligent video monitoring system which is considered to be the extension of human vision, has got considerable development. Monitoring system at present is widely used in business application , but it can't monitor the environment automatically , because they usually write down output results of the camera, after the unusual situation (if the vehicle in the parking area is stolen ) happens, security personnel could observe the fact taking place through the result of writing down , but it's too late, and we should need monitoring system one hour real-time monitoring, and analyzing the picture data catched from the camera automatically. When exception happens, the system can send out the alarm accurately and promptly to the guards, thus avoid the emergence of the crime, and reduce the cost of the manpower, material resources and financial resources at the same time. This thesis analyses the deficiency of the present intellectual monitoring system in application, has proposed improved algorithm in video capturing, motion detection, and object classification.This thesis has introduced the concept of the intelligent monitoring system and current situation of relevant research at first, and provides the prospect of its application. Then ,the thesis gives an introduction to the system structure frame of HW-S600 intelligent monitoring system , and has analyzed the relevant function of each module , has compared with and introduced the advantage and disadvantage of the algorithm that this thesis selects for use.The second part introduces the technology of videocapturing. Video capturing is the most basic part of a intelligent monitoring system, which offers the reliable video stream for following processing. The video capturing mainly depends on video capture chip. According to the different demands of performance of the hardware, our system mainly chooses two kinds of chips which are VC0323 and Hercules, to meet different demands. In the part of the software development, we adopt DirectShow technology to develop. DirectShow is a kind of software development kit, which is provided by Microsoft for stream media processing in windows platform. And it's released with DirectX together. Our system uses the latest edition 9.0 of DirectX for development. DirectShow strongly supports multimedia stream capture and playback. Using DirectShow , we can very conveniently catch the data from the video capture chip, which supports WDM driver model. And we can also process and store the video stream. It widely supports various kinds of media forms, including Asf, Mpeg, Avi, DV, Mp3, Wave, etc., makes the playback of the multimedia data easily. Meanwhile, what DirectShow is offered is a kind of open development environment, where we can customize one's own component according to people need. All the modules of the intelligent monitoring system are into this kind of component to finish their task.The third part has mainly discussed the relevant algorithm of motion detection. At first , our thesis introduces the usual algorithm in motion detection at present , including frame difference , background subtraction and optical flow. And analyses the advantage and disadvantage of the three kinds of algorithm, has mainly studied the background subtraction algorithm. In the design of the background subtraction algorithm, this thesis adopt adaptive mixture Gaussian distribution model to implement the background modeling. Meanwhile, our algorithm sets up active level for each background pixel model to reflect present change intensity of current scene. According to the change condition, our algorithm can adjust the learning rate and the sample frequency of the background model. Becauseof this feature , our background model has better adaptability to the changes of the illumination and background.The fourth part mainly discusses the algorithm of connected component labeling. Labeling connected component in a picture is to get the feature of moving object, which can be used for object classification and identification. Because traditional labeling algorithm needs to scan a binary image for two times. It is in a situation that probability happens higher in the repeated labeling. The efficiency of tradition algorithm is very low. For the requirement of real-time performance, our algorithm adopts one time scan to finish all the region labeling. Our algorithm makes use of border track technology to detect the inside border and the outside border. Meanwhile, labels all the pixels in the current component. The entire task can be finished in a one-time scanning procedure without labeling the pixels for a second time. The complexity of the algorithm is linear, which depends on the size of the frame in video stream.The fifth part mainly discussed the algorithm of object classification. Firstly, we introduce two kinds of commonly used classification methods. One kind of classification is based on shape information, and the second one is based on movement characteristic. Our algorithm makes use of shape information to classify objects. The two steps in our algorithm are following. (1) Get the attributes of the object such as width/height, size, and barycenter through connected component labeling algorithm. (2) Make use of BP neural network to classify objects. BP algorithm is a kind of effective classification algorithm, but it's low efficiency. Our algorithm has put into the momentum gradient and dropped algorithms in the procedure of training of BP neural network, which makes that neural network not merely considers the function of the error on the gradient while revising its weight and threshold value, and considers the influence of the variation tendency on the error curved surface, it allows to neglect the small change on the network.This can effectively restrain network from sinking into some extremely small point, and contribute to reducing the training time.This thesis analyses the advantage and disadvantage of each module of intelligent surveillance system finally, and put forward suggestion on how to improve the present algorithm. Meanwhile make s discussion to the intelligent monitoring system in the technological development trend in the future. The intelligent monitoring system, which based on DirectShow technology, can quickly detect moving object, get attribute of the object and classify it. And it can choose different video capture device of different performance to get video stream of different quality. The user can store and preview the captured video. Our system has offered strong support for the research in the intelligent monitoring system of Han Wang Company. And our production can be added to their algorithm library for future use.