Research On Moving Object Dectiontechnology In Surveillance Video

Traditional video surveillance system can only store historical videos for investigation after accidents, which can not timely notice and eliminate accidents. Target detection technology can solve these problems. Good detection results are related with the detection algorithm or actual environment, which is the basis of the following target tracking, feature extraction, classification, behavior understanding. The effectiveness of results directly determines the efficiency and accuracy of these studies. In real situations, the video shooted is susceptible to a various impacts of the environment, and people put forward higher requirements on the system stability, robustness, real-time ability, so video-based moving object detection technology has many challenges. To solve these problems, this paper will study the target detection algorithm.In this paper, we deeply study the traditional Gaussian Mixture Model, compared with single Gaussian model, the method describes pixel changes with more than one Gaussian distribution, which can adapt to complex environments. However, there are some flaws in GMM algorithm, the paper raised the following two questions: first, when the foreground object keeps still in a period of time, the system may be unable to detect the targe because of the constant background updation; the second is how to achieve a strategy to adaptively decide the number of Gaussian distribution in order to increase real time ability. For the first question, we compare the object’s overlapping rate of two consecutive frames to determine whether the target is moving or not, then to select the specific update policy to solve the problem; for the second problem, an adaptation strategy is put forward, and by determining the eliminated Gaussian distribution, the algorithm’s real time ability is improved. The paper also made a comparison test to the traditional GMM algorithm and proved the effectiveness and real-time ability of the improved algorithm.This paper also modified the kernel density estimation method based on key frames, which does not require prior knowledge of the distribution of pixel changes and there is no need for complex parameter settings. In the traditional kernel density estimation method, all samples are indiscriminately used to estimate the probability. The proposed keyframe method just selects key frames as samples from the sample set to represent key information of background,which greatly reduce the amount of computation and improve real-time performance. The improved method is also compared to the GMM algorithm and the traditional kernel density estimation method to prove the new algorithm’s effectiveness and real-time ability. At last, we utilize the combined detection method which integrates the modified GMM and the kernel density estimation method, and received better results, the paper also use a shadow detection strategy to recognize and remove shadows in the detected objects in the test results to avoid the interference.