Efficient generalized hough transform algorithms for modern applications

Abstract

In this research, some novel modifications and modern applications of the Hough transform algorithms have been pursued. First, the dual points Hough transform algorithm was studied. In the original approach, a point pair with the same gradient direction is selected as the features in the transform. However, it has been proven that such an approach is not suitable for some arbitrary shapes, especially for the shape containing straight-line primitives. A number of spurious votes are generated. Therefore, a modified approach is proposed to deal with this problem. Our approach optimises the R-Table of the prototype by using some statistics techniques in order to reduce the number of entries per index in the R-Table and make the entries inside the R-Table distribute evenly. Experimental results show that our method can both speed up the process and increase the recognition accuracy. Second, a dominant point detection algorithm has also been studied. The reason for pursuing this study is that we can reduce the number of operations by employing dominant points as the features in the Hough transform process. Owing to the lack of appropriate dominant point detection algorithms for real scenes, a new dominant point detection algorithm has been proposed. Apart from the curvature points, we also defined the termination and intercept points on a digital curve as the dominant points since they indeed give useful information about the natures of the curves. A weighted mask is proposed for the initial detection. By using a look up table, all possible dominant points can be located efficiently. Then an iterative suppressing is applied in order to reduce the number of dominant points. We have also designed a new suppression criterion that makes use of the average cosine angle on every boundary point to measure its significance. We suppress the dominant points when they are insignificant. Experimental results show that our method achieves a better performance in terms of approximation error when compared with other methods. Besides, our method can be applied directly to the detection of dominant points in a real scene. In addition, our method is non-parametric. After the development of the dominant point detection algorithm, we have developed a Hough transform algorithm that makes use of the dominant points as the corresponding features in tracking an object in a video sequence. The user inputs a user-defined object in the first frame. Then, an R-Table is built based upon the dominant points of the selected object. When the next frame is reached, the transformation parameters of the object can then be detected by using the Hough transform. After the transformation parameters are recovered, the algorithm will determine whether it is required to update the model by examining the length of the object boundaries and the centroid of the object. The algorithm will repeatedly track the object until the object is disappeared in the video sequence. Our proposed algorithm is capable not only to track the object under a continuous deformation, but also able to recognise the object under occlusion and recover the tracking when the object reappears in the video sequence. Finally, a Hough transform algorithm that makes use of the color information as the corresponding features is studied. To start with, a color image is segmented into some regions with homogeneous colors by applying a watershed algorithm and a new region merging algorithm. The region merging algorithm merges regions based on the ideas of reducing the errors of the color differences, maintaining the uniformity of the color and controlling the minimum size of each region. After the image is segmented, region pairs are selected to form an index inside the R-Table. The matching criterion relies on the color similarity between the regions taken from the R-Table and the target object. Our proposed algorithm can deal with a change in illumination, the presence of occlusion and similar transformation. In short, through these studies, some Hough transform algorithms have been greatly improved, in particular in terms of the speed of the process. Moreover, our developed algorithms have been applied successfully to many image recognition problems and video tacking applications, which reflect the usefulness of this study.