Research On Intelligent Scene Surveillance System And Its Application In Indoor Surveillance

Along with the development of electronic technology, communications technology and Computer Vision technology, the research and application of Intelligent Video Surveillance System are increasingly becoming an important part in human life. Video surveillance is usually carried out in some scene, so there are different detective methods according to different scenes. Scene Surveillance has already become a significant application field in modern IT technology, for it not only has wide application foreground in everyday life and social work, but also plays an important part in fields such as national defense, traffic, home, etc.As the first step of video processing in Scene Surveillance System, the technology of motion detection takes up a great important position. In addition, it is also a pop part and difficulty technically. The kennel technology in Scene Surveillance Technology is to detect, segment and identify the motive objects in the scene, which is also a very active research direction in Computer Vision domain and possesses important research value in both science and engineering. So far, motion detection technology is widely used in many fields like Intelligent Surveillance, Intelligent Robot, Alternation between Man and Machine, Virtual Reality, Medical Diagnose, Video Compression. It has become a key algorithm in security surveillance abnormal alarming tasks based on video, especially in security surveillance application. However, in actual scene surveillance, the surveillance background is always complicated, such as the shake of branches and leaves and the change of lighting brightness. These phenomenon bring huge effect to motion detection, so it is a difficult problem to detect motive objects in complicated scene. This paper is mainly doing research and discussion aiming at the problem of indoor scene surveillance and the key technology that is concerned.According to Video Surveillance System, this paper intends to recognize the unusual abnormal phenomenon automatically and send out alarm signal by using method of digital image processing and recognizing. To apply digital video surveillance system to doing motion detection and recognize operations with image sequences, make alarm signal to notice supervisors who can do measure immediately once detecting motive objects, can lighten the supervisors'vision burden in a deep degree. Besides, most digital surveillance systems contain memory module, which can store surveillance scene image data continuously. However, the long work time of digital surveillance system and huge data storage raise higher request to memory capability. Almost, the goal of storing surveillance image is to record the actions of surveillance scene. However, if we continue recording images without any motion, we lose the purpose of storing them. In this case, we also need motion detection algorithm to estimate whether there are motive activities in whole or some part of the surveillance image. Thus it can be seen that in digital video surveillance system, the technology of detecting and recognizing of motive image not only can replace some parts of supervisors'work, raise the robotic level of surveillance system, but also enhance efficiency of surveillance storage.First of all, this paper expounds the development course and status, technology problem, application range and development trend of video surveillance system, and the theory background of intelligent video surveillance system, research status and developments.In aspect of motion detection, this paper introduces three common motion object detection algorithms: light flow, background subtraction and border frame subtraction. Firstly, we discuss on the principles and available situation of each method, experiment on the last two algorithms in order to make analysis of experimental results by comparing the advantages and disadvantages of each. On this basis we bring forward a model based on adaptive background moving object detection method. This method establishes Gaussian distribution background model by taking the brightness and chroma of each pixel. The newly observation values are used to update background model adaptively, so that the system can effectively adapt to the transformation of scenes, and increase detection accuracy through shadow detection and noise treatment. We make progress with traditional two frame difference method which always brings with background miscalculation, put forward method which compare with the recent three frame images. Experiments show that three difference methods can significantly improve the detection accuracy. On the research of adaptive algorithm about the background model, this paper discusses the application of simple background (such as underground parking lots, banks, etc.) as well as Single-Gaussian background model, which is applied to complex background (such as parks, crossroads, and so on) Multi-Gaussian background model, and than analyzes the problem with the update of background model, discusses on two basic principles that should be followed when updating background model.We can distinguish foreground spots out of background ones by applying the aforementioned motion detection method, but these are all partition based on a certain threshold T as the value for the delineation of boundaries, the selection of T directly affects the quality of foreground spots, thus this paper discusses on the method of threshold selection, puts forward threshold segmentation method applicable to indoor surveillance based on Single-Gaussian background model. In order to segment moving objects from foreground spots set integrally, so that we can get the description of the characteristics of foreground goals. According to the continuity on space of objects, we utilize mathematical morphology process and connectivity regional detection algorithm to remove noise and fill the holes. Combined with the characteristics that Single-Gaussian model can obtain the estimated background images, under the precondition of reducing noise and maintain integrity objects, we analyze reasonable threshold of foreground spots after background subtracting through experiment, as a result, we can get perfect binary image of object background. In the identification of human body, we extract the outline of the foreground region; analyze the size and ratio of width and height. Basis on this, we determine whether the foreground region is human body and make security alarm if necessary.Finally, this paper focuses on discussion on detect problems that usually happen in indoor surveillance. We bring forward the unusual event detection and classification aiming at indoor surveillance characteristics. Switch lights events detection and analyze with objects moving distance were discussed separately. These are all frequently encountered problems in indoor surveillance. Based on the previous algorithms, we initially construct a system suitable for indoor scenes surveillance, which combines adaptive background model and improved adjacent frame difference method. Experimental results show that the system is not only effective in resolving the background miscalculation issue, but also can quickly adapt to the update of background. So that we can detect motive objects timely and accurately and determine whether there are human bodies.All the algorithms involved in this paper have been validated by experiments, with good results. These algorithms have good modular structure and they are easy to expand and transplant which would be convenient for applications in various scenes surveillance systems.