Object Tracking Against Wide Baseline

In the field of computer vision, wide baseline refers to a number of interference factors that cause objects'appearances to change greatly. Under wide baseline conditions, an object or scene could present different appearances, which will influence the effective representation for objects and the performance of object localization and tracking. Object tracking against wide baseline is one of issues in the fields of computer vision and machine intelligence, and the bottleneck in many related applications. Therefore, object tracking against wide baseline has great theoretical and practical significance.Focusing on object tracking against wide baseline and its key technologies, research work in this paper is carried out as following four aspects:â‘ Wide baseline is a special and real case application condition, which could influence the performance of computer vision system. This paper analyzes object tracking against wide baseline and its present, and summarizes its bugs. Then, following its essence, object tracking is conceived as a process of feature matching.â‘¡Following its essence, object tracking is conceived as a process of local invariant feature matching. Then, the experimental results show that the novel concept of object tracking against wide baseline provides great performance and real-time capability. Currently,most of the evaluation criterions as ROC etc. only concern"match or mismatch", but don't attach importance to"how many mismatches". This paper adopts reconstruction similarity (RS) which is proposed by our laboratory to evaluate comprehensively the performance of the method of feature matching.â‘¢The method of feature point matching could reveal itself during non-uniform or overly concentrate results, and then could lead to object tracking be into a lower accuracy. Thus, the paper proposes a hybrid scheme of object tracking against wide baseline with two-stage, which includes that lightweight local invariant feature matching and simplified object tracking. The both stages achieve initializing and searching precisely target. â‘£Currently, most of the algorithms of local invariant feature matching could get into trouble against greatly different views or non-grid deformation. Therefore, against non-grid deformation, the paper proposes local invariant feature aroused by the algorithm of Geodesic Intensity Histogram (GIH) feature descriptor, which embodies the method of dealing with deformation to obtain isotropy feature area and Anisotropic Geodesic-Intensity Histogram (AGIH) feature descriptor.