Investigation On Registration Of Multi-Source Dynamic Video Sequences

Image registration is first proposed in the research of auxiliary navigation system for aircraft, projection system of weapons in American in 1970s. With the development of science and technology, image registration has become a very important technique in the information technology, and has been widely used, such as computer version and pattern recognition, medical image analysis, remote image analysis, and military purposes. Image registration is to get the correspondence between the images of the same scene taken from different view points, at different time, or with different kinds of cameras.The purpose of image-to-image registration is to get the spatial transformation between the two images. However, the temporal correspondence is also required for some cases in the practical applications. Taking target identification for multi-sensor as an example, the sequences are taken from different sensors, so it is necessary to register them before multi-sensor information fusion. There are two main motivations to use sequence-to-sequence registration: (1) it can resolve spatial ambiguities and handle situations that image-to-image alignment can not:①resolving the spatial ambiguities,②improving accuracy of registration,③reducing combinatorial combinatory; (2) it gives rise to new video applications that are very difficult for image-to-image registration:①providing super-resolution both in time and space,②resolving the image registration for multi-sensor easier using the temporal behaviors in the sequence. This paper reviews the previous achievements in sequence-to-sequence registration, including feature-based, direct-based, and reference-frame-based sequences registration. And much research has been done on how to resolve the correspondence in time and the difficulties caused by multi-sensor. The paper proposes and realizes the registration methods addressing visible light and infra-red sequences, and the experiments show the validity and the accuracy of them. The main research content of the paper is as follow:1. Analyzed and summarized the previous methods, and improved the methods based on high-pass energy transformation to address the abrupt changes caused by different sensors, which satisfied the acquirement of accuracy and real-time and was used to register the corresponding frames in time domain.2. Analyzed the registration methods based on matching contours, and realized the matching method using improved 8-d chain codes. And it provided the theoretical support for matching the trajectories.3. Researched on the sequence-to-sequence registration based on scale change of target. In the method, we acquired the temporal transformation according to the change curves of target's scale, and used the method based on high-pass energy transformation to register corresponding frames to get the spatial transformation parameters.4. Researched on the sequence-to-sequence registration based on trajectories. In the method, we represented the trajectory in the form of contour, used the contour-matching method to match the trajectories, and finally acquired the correspondence of the two sequences both in time and space.