The Research On Objects Tracking In Image Sequences

The work described in this dissertation is supported by Aeronautical Foundation of China (NO. 02153073) and the project of ground target tracking in complex background. Object tracking through image sequences has been an active field in computer vision, image processing and pattern recognition. Tracking is carried out by analyzing the image sequence from the sensor, identifying the independently moving regions or those that users are interested in, and locating them in the image. It's of great importance not only because it provides the track of the object, but also because it offers reliable information for object recognition, scene analysis and other applications. As a topic with many applications, object tracking has drawn much attention of researchers. Many institutes have done a lot of work on it and got achievements. However, there are still many problems to solve in order to build a robust and practical tracking system.In this paper we focused on objects with a certain area, and aimed to improve the precision and robustness of the tracking algorithm. My research work can be summarized as the following aspects:1. The well-known mean shift tracking algorithm developed by Comaniciu describes a target just by three degrees of freedom, and cannot cope with the rotating targets. Aiming at this shortcoming, we extend the kernel radius parameter to a diagonal matrix and propose a bandwidth matrix mean shift method. We also prove its convergence. Applying the proposed method to image tracking leads to a new 5D tracking algorithm that describes targets by five degrees of freedom. The new algorithm adapts to the complex movements of targets, including rotation and width to height ratio variation, and solves the problem of rotating targets.2. We propose a Gabor wavelet based feature point detection method, which is suitable for image stabilization. It effectively eliminates the wrong pair of feature points by one-to-one mapping. Experiments show that the detected feature points can describe the precise displacement between two adjacent frames, thus form a reliable base for the image stabilization.3. Current feature points based image stabilization methods often fail on the low S/N ratio image sequences in which there is a large translation of the background. Aiming at this problem, we propose a new stabilization algorithm based on the trust region method. It can cope with the abovementioned image sequences satisfactorily and has the advantage of low computational burden and high precision.4. The mean shift tracking algorithm developed by Comaniciu is deficient in describing targets by discrete parameters. Aiming at this deficiency, we combine the orientational multi-scale normalized Laplacian filter and the trust region method, and propose a new trust region scale-space target tracking algorithm. It can describe the target's rotation and scaling by continuous parameters with high precision.5. We propose a feature selection method based on the continuous feature space. Compared with the discrete feature set based method, it can choose better feature for tracking with less number of features evaluated. Integrating this method into the trust region scale-space tracking algorithm leads to a better one which can cope with the change of a target's color and brightness. The new algorithm performs well on image sequences on which the mean shift algorithm fails.6. For low S/N ratio image sequences on fast moving platforms, we propose an effective trust region Kalman filtration based tracking algorithm. It employs the trust region based stabilization method to eliminate the influence of platform's movement, and the Kalman filtration to eliminate the image noise. It enlarges the search region and reduces the number of template matching operation as well.