Study On The Method Of Moving Target Detection And Tracking In Image Sequences

The moving target detection and tracking in the image sequences is a hot research topic in computer vision field. Many people have paid attention to it. Some key problems of the moving target detection and tracking in the image sequences are discussed in this dissertation Such as the correctness, real-time performance and anti-jamming of moving target detection algorithm; shadow detection and occlusion handling algorithm; the non-linearity moving targets tracking based on particle filter; vehicle targets tracking based on moving camera; the license plate detection during vehicle targets tracking; license plate image pretreatment, character segmentation, extracting character feature and character recognizing. The main research results are as follows:By researching moving targets detection algorithm based on background model, a moving targets detection algorithm including background building model based on Median filter and background updating model based on Wiener filter is proposed. The reference background image which is setup by Median filter provides a fine initial state for background updating and moving targets detection.So the moving detection effect in the initial time is improved. Furthermore, the time for updating initial background image is decreased. The proposed background updating model restrains the main interfere in the complex scene by the linearity prediction function of Wiener filter. Because of the fast calculation speed and optimal estimate with minimum variance of Wiener filter, the moving targets detection during background updating can quickly and effectively restrain the dynamic noise in the scene. The effect and real-time performance of moving targets detection is good.By researching moving targets tracking algorithm based on fixed camera, a multi-targets tracking algorithm based on Kalman filter and color probability histogram is proposed. During the targets are tracked, three kinds of feature is used. The features respective are the position featue and shape feature based on Kalman filter, the color feature based on color probability histogram. And tracking match function is proposed to ensure the accuracy of matching target during tracking. Moreover, the searching area is reduced by the prediction information of Kalman filter, so the calculation speed of whole tracking algorithm is improved. At last the multi-vehicle targets are successfully tracked by the proposed algorithm.By researching shadow detection and occlusion handling algorithm, an interactive image segmentation algorithm based on Kalman filter is proposed for resolving the problem of shadow and occlusion. And the synchronous handling the shadow and occlusion is realized. The mutually influence between shadow detection and occlusion algorithm is avoided. Random walk algorithm based on Kalman filter can extract the seed points without human operation. At the same time the calculation region of random walk algorithm is reduced by the prediction of Kalman filter. The random walk algorithm without human operation is realized.The linearity and non-linearity state of moving targets is researched. The linearity state estimate modificatory formula derivation is implemented based on pinhole imaging and geometry theory. Then the modificatory formula together with Kalman filter is used to estimate the linearity state of non-linearity moving targets. The estimate results of the proposed method are better than only Kalman filter’s estimate results. Furthermore a improved variable rate particle filter is proposed, according to the charecteristic of non-linearity state. The measurement model of improved variable rate particle filter is simpler than traditional variable rate particle filter. And the target moving angle changing value of tracked targets is brought into the adjustive strategy of state estimate frequency. The state estimate precision when the targets trajectory is curving or sudden changing is improved with less calculation time.For solving the practice problem of tracking and recognizing traffic targets, an application system frame which involves the tracking moving vehicle targets based on moving camera, license plate real-time detection and recognizing is proposed. Multi-resolution Lucas-Kanade sparse optical flow based on wavelet pyramid is proposed to solve the traditional Lucas-Kanade optical flow error matching target during tracking rapidly moving targets.The license plate detection algorithm based on plumb edge is proposed, which can detect license plate during tracking vehicle targets with real-time performance and establish a foundation to following license plate character recognizing. Aim at the characteristic of license plate character which is shot during tracking the moving vehicle, the license plate image pretreatment and character recognizing algorithm are proposed. At last, the license plate character recognizing results is satisfying.