Algorithms Study On Smart Visual Surveillance System

In the modern time of pursuing for the high automatization in the field of visual surveillance, the desire for a smart surveillance system, which is the substitute of surveillance members and can monitoring the moving object(such as human, cars and so on) in some surveillance scape automatically and consecutively, is very strong. In most cases, this intelligent system has some functions such as motion detection and object tracking etc. It's very possible to design this system which no one needed to watch by full utilizing the powerful calculating capacity of modern computer and many subjects such as image processing and analyzing, computer vision etc. But regrettably this kind of system is not appeared in our country currently. The most difficult problem is lack of a device of global algorithm. Common algorithm often cannot reach the practical standards such as tight linking among modules, low algorithm complexity, low ratio of missing alarm and false alarm etc. This paper is put forward in such a background. The algorithm complexity is more lowered by simplifying and improving the algorithm on the two modules of motion detection and object tracking on the basis of researching and analyzing many other person's algorithms and studying others' research fruits. The global algorithm synthetically considers the real-time property and robust performance, and it can be used in surveillance system.Firstly this paper introduces the research status and tendency in the world in detail. The whole paper is divided by two modules of motion detection and object tracking based on its function, and the article elaborates them one by one.Motion detection module is the most important part in the intelligent system. Its device is the key of the whole system and it also holds the most calculating task. This module is divided by three sub-modules of pre-processing, motion detection and post-processing. In the sub-module of pre-processing, the paper analyzes performance of traditional filters and points out its advantages and disadvantages, then presents a modified multilevel median filter by importing the idea of weight and classification. The proposed filter can not only suppress noise more effectively but also preserve the features of images. In the sub-module of motion detection, the article first analyzes performanceof traditional algorithm and points out some disadvantages, and then brings forward one improved algorithms: the modified dynamic background deduction algorithm, which puts the advantages of temporal differencing algorithm and dynamic background deduction algorithm together to present a dynamic background model. Experimental results are given to demonstrate its validity and practicality. The post-processing sub-module uses morphological filter to remove isolated noise and to fill the holes inside objects, and the traditional method of this sub-module is improved in the paper.Object tracking module is also an important part in the intelligent system. Among them the object tracking with an Active Camera has much more practical value. Firstly, the article introduces the traditional matching tracking algorithm detailedly in the fixed background. Secondly, a simple tracking algorithm integrated with many moving objects' characters such as smooth motion property and consecutive size property is brought forward. The simplified algorithm is just tracking the outer-rectangle of moving object area in-the binary image. Carefully considering the problems of the occlusion, temporary disappearance and being permanently moved in or moved out of objects detailedly, this algorithm can be used in the tracking of multi-object and its robust performance is also strong. Finally, the paper analyzes the object tracking with an active camera.At the end of this paper, some demo sequences of experiment are given and elaborated. Finally the paper introduces shortcomings of proposed algorithms and aspects deserved to be strengthened.