Research On Deformed Target Locate Algorithm Based On Extended Contour Description

Object detection is one of important computer vision research field. A lot of research work has been carried out in the past few decades. But it's also difficult for deformable target detection. According to some new thoughts, this paper puts forward some new object detection thoughts to solve some problelms during the process of object detection.Classic generalized Hough transform can locate non-deformed shape object, while it is difficult to solve the problem when the target is deformable. A two level deformed target locate algorithm based on variant of the well-known generalized Hough transform for solving this problem is presented in this paper. Firstly a two-level location structure from coarse to fine strategy is introduced to reduce search range from whole image space, in coarse location step, the edge local binary pattern histogram features are extracted to detect the range of the target. In fine location step, a large dispersion window is presented to merge voting results. The experiment results demonstrate that the method is effective to the deformed target locating while the time and memory cost is much less.Generalized Hough transform asks good contour detection. And its single template type impactes the target locacting accuracy. This paper puts forward an object detection algorithm based on codebook in angle spread Shape Context descriptor aiming this problem. Classic Shape Context is a good shape descriptor. A common problem for the classic Shape Context is that when two contours are in different angular bins, similar contours have very different histograms. The way we overcome this problem is to overlap spans of adjacent angular bins. Experiments show that our method is capable of detecting object effectively and is valuable in computer vision field's applications. An algorithm of generalized Hough transform to locate the object introduced by weighted key feature points is proposed in this paper aiming at the problem that generalized Hough transform is always wrong when the target is slight deformed or interfered by noise. Given that the key feature points posses steady image information, the key feature points should be weighted to outstand the key feature points information. Vote coefficient can be constructed based on this. It makes detecting peak of accumulator easy and fault less. Then we can get the final result by the peak of accumulator.