شناسایی سریع هدف با استفاده از برنامه ریزی پویا از ترکیب گروه های خط برجسته

This paper presents a new method of grouping and matching line segments to recognize objects. We propose a dynamic programming-based formulation extracting salient line patterns by defining a robust and stable geometric representation that is based on perceptual organizations. As the endpoint proximity, we detect several junctions from image lines. We then search for junction groups by using the collinear constraint between the junctions. Junction groups similar to the model are searched in the scene, based on a local comparison. A DP-based search algorithm reduces the time complexity for the search of the model lines in the scene. The system is able to find reasonable line groups in a short time.

This paper describes an algorithm that robustly locates collections of salient line segments in an image. In computer vision and related applications, we often wish to find objects based on stored models from an image containing objects of interest [1], [2], [3], [4] and [5]. To achieve this, a model-based object recognition system first extracts sets of features from the scene and the model, and then it looks for matches between members of the respective sets. The hypothesized matches are then verified and possibly extended to be useful in various applications. Verification can be accomplished by hypothesizing enough matches to constrain the geometrical transformation from a 3-D model to a 2-D image under perspective projection. We first extract junctions formed by two lines in the input image, and then find an optimal relation between the extracted junctions, by comparing them with previously constructed model relations. The relation between the junctions is described by a collinear constraint and parallelism can be also imposed. Junction detection acts as a line filter to extract salient line groups in the input image and then the relations between the extracted groups are searched to form a more complex group in an energy minimization framework. The method is successfully applied to images with some deformation and broken lines. Since the system could define a topological relation that is invariant to viewpoint variations, it is possible to extract enough lines to guide 2-D or 3-D object recognition. Conventionally, the DP-based algorithm as a search tool is an optimization technique for the problems, where not all variables are inter-related simultaneously [6], [7] and [8]. In the case of an inhomogeneous problem such as object recognition, related contextual dependency for all the model features always exists [9]. Therefore, the DP optimization would not give the true minimum. On the other hand, the DP method has an advantage in greatly reducing the time complexity for a candidate search, based on the local similarity. Silhouette or boundary-matching problems that satisfy the locality constraint can be solved by DP-based methods using local comparison of the shapes. In these approaches, both the model and matched scene have sequentially connected form of lines, ordered pixels, or chained points [10], [11] and [12]. In some cases, there also exist many vision problems, in which the ordering or local neighborhood cannot be easily defined. For example, definition of a meaningful line connection in noisy lines is not easy, because the object boundary extraction for an outdoor scene is itself a formidable job as object segmentation. In this paper, we do not assume known boundary lines or junctions but open any connection possibilities for arbitrary junction groups in DP-based search. That is, the given problem is a local comparison between pre-defined and sequentially linked model junctions and all possible scene lines in an energy-minimization framework. Section 2 introduces the previous research about feature grouping in object recognition. Section 3 explains a quality measure to detect two line junctions in an input image. Section 4 describes a combination model to form local line groups as well as how junctions are linked with each other. Section 5 explains how related junctions are searched to form the salient line groups in DP-based search framework. Section 6 gives a criterion to test the collinearity between lines. Section 7 tests the robustness of the junction detection algorithm by counting the number of detected junctions as a function of the junction quality and whether a prominent junction from a single object is extracted under an experimentally decided quality threshold. Section 8 presents the results of experiments using synthetic and real images. Finally, Section 9 summarizes the results.

In this paper, a fast and reliable matching and grouping method by the dynamic programming is proposed to extract collections of salient line segments. We consider the classical dynamic programming as an optimization technique for the geometric matching and grouping problems. First, the importance of grouping for object recognition is emphasized. By grouping together line features that are likely to have been produced by a single object, it has been well known that significant speed-ups in a recognition system can be achieved, compared to performing a random search. The general fact is used as a motive to develop a new feature-grouping method. We introduce a general way of representing line patterns and of using the patterns to consistently match 2-D and 3-D objects. Main contribution in this paper is a DP-based formulation for matching and grouping of line patterns by introducing a robust and stable geometric representation that is based on the perceptual organizations. The endpoint proximity and collinearity consisting of the image lines are introduced as two main perceptual organizing groups to start the object matching or recognition. We detect the junctions as the endpoint proximity for the grouping of line segments. Then, we search again a junction group, in which each junction is combined by the collinear constraint between them. These local primitives, by including Lowe's perceptual organizations and acting as the search node, are consistently in a linked form in the DP-based search structure. We could also impose several constraints, such as parallelism, the same line condition, and rotational direction, to increasingly narrow down the search space for possible objects and its poses. The model description is pre-defined for comparison with the scene relation. The collinear constraint acts to combine the two junctions as a neighborhood for each other. The DP-based search algorithm reduces the time complexity for the search of the model chain in the scene. Through experiments using images from cluttered scenes including outdoor environments, we demonstrate that the method can be applied to optimal matching, grouping of line segments, and 2-D/3-D recognition problems, with a simple shape description that is sequentially represented.