Research On Dynamic Sence Modeling And Object Tracking For Intelligent Surveillance

As an emerging research field, the intelligent video surveillance technology is a hot topic of computer vision in recent years. The intelligent video surveillance, includes several key factors such as video super-resolution reconstruction (SR), object segmentation, scene modeling and target tracking, and these factors form the basis of high level video analysis, inlucding target recognition, behavior analysis, etc. Based on the video analysis and application, this paper carries out the deep research on the intelligent video surveillance and related technologies.Firstly, video SR algorithm is developed. In the SR algorithm, a key step is the video motion estimation. Compared with other methods, Matching algorithm based on features of video has higher robustness. However, the accuracy of this kind of methods is affected by the position and selection of feature points. To overcome this problem, this paper introduces the particle filter into the motion estimation to reduce the matching error. The main disadvantage of the particle filter is particle degeneracy. In this paper, an extended kalman filter is used to general the proposal distribution, and an unscented kalman filter is used to refine particles. The experimental results show that, compared with other eight classic filtering algorithms, the proposed algorithm has much better performance; and for the SR issue, the proposed algorithm can estimate the camera motion more accurately.Secondly, image threshold algorithm is developed and a novel particle swarm optimization (PSO) algorithm is proposed. OTSU algorithm is commonly used in image threshold segmentation algorithm; but it can not complete multi-threshold segmentation tasks due to the huge calculation. To address this issue, firstly this paper presents a new PSO method-local mutation particle swarm algorithm. The experiments show that compared with currently popular PSO algorithms, the proposed algorithm can achieve better results, and can effectively address the OTSU multi-threshold segmentation problem.Thirdly, in the dynamic scene modeling step, an adaptive Gaussian mixture modeling method is proposed. Gaussian mixture modeling method is widely used in background modeling method. However, the updating rate of the background remains a difficult issue. Currently, all regions use a fix updating rate in Gaussian mixture modeling. To overcome this issue, this paper proposed an adaptive Gaussian mixture modeling method. In the proposed method, a general Gaussian mixture model and a frame difference method are combined, and the key target of this method is that different updating rates are used in the foreground and the background. Meanwhile, an adaptive updating operation is used in the foreground region. Experimental results show that, compared with some typical background modeling methods, including the frame differencing, the proposed algorithm is able to filter out local motion within the scene more effectively, and is able to eliminate background noise and the shadow of objects.Fourthly, based on the analysis of Camshift, a fast Camshift algorithm(FCshift) is proposed. Camshift is a widely used in target tracking method in video sequences, however, sometimes it requires high level real-time feature. So how to reduce the computational load is still a difficult issue, and there is few research about it at present. Based on modeling the iteration process and the computation load of Camshift, this paper deduces the relationship between the size of the extended window and the computational load of Camshift, and then proposes a novel fast Camshift-Fcshift. The experiments show that, in the same scene, compared with the basic Camshift algorithm, the running time of the proposed algorithm is reduced over20%.