| Intelligent video surveillance system could carry out automatic analysis of camera image sequence,dynamic detection and tracking of moving targets in the scene using computer vision and image processing methods without human intervention.The behavior of the target could be judged,and the abnormal in the scene could be identified and reported to the police.With the development of digital video technology,this new research subject has become more and more important,which has an extensive application value and a great developing prospect in various aspects including national security and civil field.The purpose of the research is to improve the accuracy,real-time performance and adaptability in various challenging environments of moving target detection and tracking algorithm.The main research results are as follows:(1)The research on target detection has been carried out under complex environment,such as background with a lot of noise,illumination change,slow moving target and high density population.A target detection algorithm combining background difference method and three-frame difference method is proposed in view of large amount of noise in the background.After the logic 'or' of the two-valued images of these two methods,the accurate motion target contour could be obtained.A foreground segmentation algorithm based on background subtraction and image time analysis is proposed for light change and background slow moving.The average photometric method updates all the pixels in the background model,and the processing object moves slowly over a long period of time.The motion estimation between successive frames is used to solve the sudden change in luminance conditions.For detection in high density population,the gaussian mixture model and the improved adaptive gaussian mixture model combined with K distribution is proposed.In order to accomplish complex background and multiple moving targets detection under different lighting conditions,adaptive changes in weight and adaptive K choose in the number of gaussian components of each pixel are carried out.(2)Shadow problem of illumination on target is studied,and an improved shadow detection criterion based on Phong object illumination model is proposed to remove the shadow.The principle of Phong lighting model is studied and applied to grayscale image.The luminance value of pixels in the scene is analyzed,and a relative change of brightness is defined,and it is deduced that this parameter is relatively stable in the shaded region.Motion target area pixel value change is big,and the change of covariance value of corresponding covariance shadow region pixels is small,stability is measured by covariance value on a 5x5 template,and an improved shadow sentence type II is deduced.According to the relative shaded area brightness variation along with the change of time remained relatively stable,a filtering template is designed for gray image filtering,so as to increase the instability of the target area.Through the threshold comparison judgment,the improved shadow sentence type III is obtained.By using these two simplified shadow judgments,the shadow detection is carried out and the experimental results are evaluated qualitatively and quantitatively.(3)With respect to the tracking problem of fast change or occlusion of target,a Camshift tracking method based on "weight histogram" and Kalman filter is proposed.Due to the influence of the generation of target histogram on the accuracy of target tracking in continuous mean shift algorithm,a "weighted histogram" method is proposed to create the target histogram,and the HSV three-dimensional target histogram is created.In the process of tracking,considering the difference of moving targets speed,different weights have been assigned to image pixels of difference image and color probability distribution image.And this method could improve the moving target tracking accuracy to some extent.If the tracking is failed when the tracking moving target appears occlusion or under fast movement,kalman filter and mean shift algorithm are combined to forecast and track the target.If the tracking target is covered,the filtering result is used as the output of Camshift algorithm to ensure the continuity in tracking.And the robustness of the tracking algorithm is improved.(4)The real-time problem of high frame frequency target tracking,an improved FPGA algorithm based on centroid tracking and background difference tracking is proposed.The problem of low cost real-time computer vision in mobile platform is analyzed.The algorithm is simple to calculate and occupies less resources,which is suitable for implementation on FPGA.Background subtraction method is effective and computational for moving objects detection in complex environments,but the background is often changed.An improved algorithm based on background subtraction is proposed,which can update the background image adaptively and reduce the impact of background change on the detection effect.Considering the computational complexity and real-time performance of the system,an improved algorithm based on adaptive centroid window tracking is proposed to track the target.Real-time target detection and tracking system can be realized on configurable devices.Before tracking video compression coding,according to the design requirements of DCT transform and quantization module,the multiplication and accumulation division unit was designed to display the real-time motion detection and tracking results through VGA display.It has good real-time accuracy in tracking moving objects. |
PDF Full Text Request