Monocular Visual Tracking Based On Scene Understanding

Visual target tracking is an important part of the current computer vision system?With the increasing demand for visual tracking system in various fields of life and the rapid development of computer science and technology,target tracking technology has been developed rapidly,but also faced with challenge of complex background,real-time and accuracy requirements.As the visual target is always in a particular scene,the understanding of the scene is conducive to the realization of visual tracking.According to human visual experience,when humans focus on a target,is not going to understand and focus on the specific properties of objects in the scene,and visual tracking requires a fast calculation speed,the current scene understanding method is too complex,it's not suitable for practical application of visual target tracking.Therefore,it is necessary to design a scene resolution method that suitable for target tracking.Unlike conventional scenario understanding methods,this paper designs a scene analysis method suitable for visual tracking,this method ignores the specific properties of the object,interpreting the relative positional relationship of each object in the field of view(including relative depth information)in order to assist the target tracking.First of all,using threshold segmentation method with simple implementation and small computation to split the monocular scene,based on these we get the two-dimensional location information of each object,and make full use of the analytic geometry relations between the real three-dimensional scene and the two-dimensional imaging images to obtain the 2D position relation and depth information of each object.Finally,2D positional and depth information are used to deal with interference and occlusion problems,and combined with meanshift tracking algorithm to achieve a more robust target tracking.