Long Sequence Of Images Target Tracking And Implementation

The research of object tracking algorithms and its applications is one of an important branch of the computer vision. At present, it is widely used in science and technology, national defense, aerospace, medical and health, as well as all areas of the national economy. The key technicals to achieve object tracking are a better division of the target, a reasonable extraction and selection of features, as well as a precisely tracking of the object, at the same time real-time tracking should be taking into account.In this paper, the algprithm of object tracking in long sequence images is proposed, which is including three primary stages: feature extraction, point matching and template updating. The point matching algorithm and template update strategy are highlighted to discuss. The work I've done is listing as follow:(1)The joint matching of optical flow registration and color distribution matchingIn order to represent the deformation of the object more accurately, a transformation matrix which is modeled by an 8-parameter vector is employed. The correspondences between reference image and candidate image can be mapped by the transformation matrix and its inverse matrix. This mapping is employed in optical flow registration and color distribution matching, then combinate this two estimation process to update parameter by M-estimator. Optical flow registration is very accurate but needs close prior alignment; color distribution-based matching, on the other hand, is robust in displacement and deformation but has relative poor localization accuracy. The point matching algorithm which combines the advantage of both approaches is not only insensitive to large translation, rotation, scaling and deformation, but also has superior localization accuracy.(2)A novel template-update strategyIn this paper, a novel template-update strategy is proposed based on the classical approach. In view of the proposed point matching algorithm tolerate large translation and deformation and has sub-pixel precision. My thought of template updating is try to change the template as little as possible in order to prevent drift and maintain the tracking accuracy. In stead of pixel information, color distribution is employed to build the discriminant. Strategy determines whether to update the cuurent template by computing the error between object and template.(3)The software of object tracking in long sequence images algorithm is implemented.The software consists of three primary modules, they are read and display module, image pre-processing module and object tracking module. Some feature extraction methods and edge detectors are employed to help the selection of features in pre-processing module. In object tracking module, the object tracking in long sequence images algorithm is provided, and two additional tracking mothods, template matching and Mean Shift tracking, for performance testing and comparing. Chapter 4 is on this topic.Through the experiment, the proposed tracking algorithm has been tested and compared with the additional two methods. The experiments show that the proposed tracking algorithm is more accuracy, and has the advantage of achieve effective tracking in long sequence images which object generate variety of deformations, illumination changes and short-term occlusion.