Research And Implementation Of Camera Track’s Analysis Algorithm Based On Monocular Vision

At present, computer vision system is widely used in these domains that are visual inspection, visual navigation and assemble automatic. In most of visual systems, camera is the tools of shooting, and the shot objects are processed. However, as a passive sensor, if the environment becomes complex and severe, it is not suitable for human to manipulate camera, the scope of its work is largely limited. In order to realize the camera’ automation and expand its scope of work, the camera should be provided with the ability of autonomous return in various complicated working scenes. In addition, the economical expenditure should be saved, if the camera is moved without navigation facility on land. So, it is the urgent demand of realizing the camera’s automation to research the ability of autonomous return, in which how to describe the motion route of camera is one of the key tasks.In this paper, without GPS or ground guide, the camera is moving in the3D space. To realize the function of automatic return, the motion route is known in advance. It is the main work. According to the global motion which is caused by the camera’s motion, its route is restored only by using single camera. Analyzing the images without fixed object is under emphasized researches which are got in the proceeding of the camera’s motion.First the relationship of the camera’s motion and the points’motion in the image is researched. Affine transformation, one of the motion models, is selected to describe the simple motion way, such as translation and rotation. And then, the motion parameters are solved. Next, the problem how to get the information by image features is researched. The point is selected as the matching element. To select SIFT which the number of feature points is large and result is steady is as the matching method. At the end of this paper, because the number of feature points is too many, the methods of sifting and matching feature points are done deep search. To do experiment for the original method, it finds that more feature points that are computed and more exact about the precision of restored motion, but the operation time is increased along with the number of feature points. At last, the method of sifting and matching feature points based on area-degree of density is given combine edge with area features. By experiment, it is proved that the method has good effect on precision and time.