| As the rapid improvement of the hardware of the computer, intelligent video is not a rare topic anymore. You can find it in the traffic control and the secure surveillance. What's more, it is also very meaningful and significant to people's live and entertainment. There are many things we can do with the help of intelligent video. As within the traffic management or secure surveillance, people can avoid watching the video all day long instead of doing something more important. The intelligent video can help them with item identification, speed measurement or stat of vehicle's number, and even the track of the target. But, however, what talked above is just a beginning. Although it gets really great development on intelligent video, that doesn't mean it has reached perfect situation, it has its shortcoming. Moreover, there must be much more useful or interesting thing that we can do with the intelligent video. We still got a lot to do. Of course, I can't do all the things by myself. Hence, in this paper, I put all my energy on the first problem, to improve the imperfect of the intelligent video.What I did is based on Open Source Computer Vision Library (OpenCV) and Visual C++ 6.0. My work is to improve the current motion detection algorithm which based on hybrid Gaussian background modeling. In the first chapter, it's mainly about the background of my research. The main part of my paper is the second, third and fourth chapter while the third is the core part.The second chapter has two main parts. The first is the basic knowledge of OpenCV, here I give a general introduction on the basic data structure of OpenCV and some important function which deal with motion detection. The following is about my research and study on the common motion detection based on sub-background algorithm. They are frame difference algorithm, Median filter, single Gaussian model and Gaussian mixture model, compare their efficiency and find their strong points and shortcomings by experiment. Besides above, I also introduce two methods that improve the Gaussian mixture model.As I said above, the third is the core part in this paper. After detailed analyze on the Gaussian mixture model, I bring forward some improvements. And I divide them into 3 sub-parts to talk about. The first sub-part focuses on the efficiency of traditional Gaussian mixture model. Because the traditional Gaussian mixture model processes all the frames the same without any concerning about the motion detail in it. Actually, sometimes, there is no motion target in the scene, and what's more, this kind of situation may last for a while. There is no necessary to compare each pixel with every Gaussian model while the procedure is very complex. I think this work is kind of waste. Therefore, I raise a method that controlling how the system work by importing an active factor. When there exists motion target in the scene, the system process the current frame normally by hybrid Gaussian algorithm, at the same time, recording the active degree. If the active degree reaches a threshold, then the system turns to leisure mode. In the leisure mode, the system only refreshes the Gaussian model which has the highest weight. Further more, I utilize frame difference which is easier and faster to detect the motion instead of hybrid Gaussian function. As soon as frame difference detects moving target in the scene, the system turns into busy mode.The second problem I deal with is about the false target in the scene. I figured one method to detect this kind of false target. My method is composed of two steps. The first step is to find the source of false target. With acknowledge that the moving target usually appears in one side of the scene and disappear in another side (maybe the same side), that means you can find a track of it in the scene. But, it's different to the false target, because the false usually appears suddenly in the scene, there is no history track for it. Hence, according to this principle, I can detect the false target by recoding each moving target's track. If some moving target doesn't have its history track, it must be a false target. The second step is to locate which part of the false target is the exact false item. I notice that the false item stay its look since it appears in the scene (at least for a short time), so I can find which are is the false zone by compare the neighbor frame. The common part in the false target must be a part of the false item.The third problem I talk about in this paper is the shadow problem. If we use hybrid Gaussian algorithm to detect the moving target, it usually get bad results when the sun or other light source is great. The reason is the shadow of the target. Because the shadow is easy to be seen by people's eyes, and they moves along the target which makes it more difficult to get rid of it from the target. If you are careless, you may lose some important moving targets because of the shadow. To take care of the shadow problem, one method comes to me. After lots of research and experiment, I found that shadow generally adds even gray value to the background it covers. According to this discover, firstly, I can detect the whole moving target includes its shadow by using traditional algorithm. Secondly, do frame difference between current frame and the background. By the way, here we only consider the target area. It's faster in this way. After I got the difference result, it's not a complex matter to locate a average area by FindContour(). Frankly, there is still one point I have to say:the gray difference between the shadow and the background is different according to different environment, how to set this value automatically worth some extra work.You can find some introduction about Image denoising in chapter 4. In this part, basically, I research and analyze mean filter, median filter,Wiener Algorithm and Mathematics Morphology filter,find their strong point and shortcomings by doing some experiments. Through analyzing the image in motion detection, choose a proper algorithm to improve the result.Finally, I did some summarize on my whole work and displayed my own opinion on the future on motion detection. |
PDF Full Text Request