Research On The Corner Detection Technique Based On Dynamic Region Of Support

As a low-level image feature,corners have been widely used in many computer vision tasks.Existing contour-based corner detectors commonly use a fixed-radius region of support to detect and evaluate corners,and thus these methods are not adaptive to the local structure of the image.As a result,the produced detection results always have unsatisfactory qualities(e.g.,inferior recall rate and location accuracy).For that,in this paper the concept of"dynamic region of support" is first studied,and then used to conduct the dynamic region of support-based corner detection methods.The main contributions of our work is documented as follows.1)To efficiently address the above-mentioned problem,a single scale dynamic region of support corner detection algorithm(DRS)is proposed.In our DRS,with the constraint of a pre-determined error tolerance,the maximum length of the straight line segment is searched along the curve to both sides to obtain a region of support.After that,based on the obtained region of support,a calculation method of "cornerity" is developed and used to evaluate a point by calculating the possibility of the current point can be a corner.Finally,a corresponding corner detection algorithm is established.2)In order to further improve the detection performance of DRS algorithm,a multiscale dynamic region of support(MDRS)corner detection algorithm is proposed.In our MDRS,a set of scale parameters,i.e.,multi-scale tolerance errors,is first determined followed by calculating multi-scale "cornerity" at each scale by a fusion strategy.Based on the obtained multi-scale "cornerity",the possibility of the current point can be a corner is evaluated.Finally,a corresponding corner detection algorithm is implemented.3)For comprehensively and effectively evaluating the performance of corner detection algorithms,a new evaluation dataset was constructed and two evaluation metrics were proposed.Firstly,in response to the lack of publicly available datasets with ground-truth corners annotations in the field,a method of ground-truth corners annotation based on line segment detection is proposed,and the annotation of the datasets is completed accordingly.Then,the existing metrics were analyzed and two evaluation metrics,"Repeatability referenced to ground-truth corners" and "Location error referenced to ground-truth corners",were proposed.Finally,the algorithm evaluation work is completed.Extensive experimental results show that the proposed algorithm based on "dynamic region of support"(DRS and MDRS)can deliver superior performance and exhibit higher robustness over the existing state-of-the-arts.