论文部分内容阅读
Abstract: The problem of road sign detection and recognition is very important in many practical problems, above all for road cadastral authorities. In this sense, an automatic application able to identify the kind of a road sign starting from common imageries such as photos could be very helpful. The main difficulty is due to a possible poor graphical definition of the imagery. In this case, a valid support can be provided by the use of Hough Transform. This paper is an evolution of a previous article, where new perspectives and examples about the development of a GIS oriented to the road cadastre management are analyzed, in order to improve and refine the methodology for an implementation of an effective, automatic, robust and reliable decision system to support technicians. It is based on the use of the Standard Hough Transform in order to detect the shape, i.e. the macro-class, of road sign (e.g. circular, squared, triangular, etc.). Subsequently, the road sign characterization has been refined by using the generalization of the Hough Transform in order to detect the specific sign within its previously established macro-class.
The production of GIS is oriented for the management of the road cadastre, the quality and the benefits of the services furnished, which do not depend only on the ability of the software component to extrapolate and make explicit the information contained in its databases in implicit form, but above all in the updating the same data bank, especially within the cases characterized by a specific dynamism.
In fact, the elements of the territory which must be represented (the horizontal and vertical system of traffic signs as well as the advertising posters) are characterized by a remarkable variety, involving both their spatial location and the alphanumeric attributes useful in order to their specialization. On the other hand, the integration of the system with a GIS concerning other territorial aspects is made possible only starting from a continuous and constant updating of the data bank related to the road cadastre, so avoiding the incongruence of the results from such an integration [1]. Nowadays, different commercial solutions are available in order to speed up the surveying of interesting elements along the roads. In this way, it is possible to collect a cartographic dataset by exploiting suitable hardware and photogrammetric software tools. The dataset is able to show the records of interesting elements along the roads in a very quick way, if compared to the classical topographic methods [2]. Nevertheless, even if the acquisition step is completely automated, the post-processing procedure is still now affected by a massive intervention of technicians, with a resulting slowdown for the surveying and positioning. Naturally, it becomes a problem in such fields given the automatic and real-time work requisite of such applications, e.g. if it is necessary fast interventions on roadways with, at the same time, an accurate consideration of the available economic funds [3]. The proposed approach aims to further investigate
? The exportation of the acquired data towards a GIS, where the last corrections will be carried out in order to respect the topological constraints, according to the already existing cartography.
computational level the GHT roughly consists in making each examined pixel in an image to“project” a copy of the searched pattern at various angles and scales (usually, projection and comparison takes place starting from the center of a certain object, but it’s possible to start elsewhere under special conditions) and then keeping track of how many pixel matches for a given scale and angle occurred between the “projection” and the tested image. Therefore, the most general algorithm definition says to do exactly that: creating a special reference data structure (usually a binary image, in the form of a table, called R-Table) and essentially comparing its “boundary” or “contour” with groups of pixels (having a central or boundary pixel for reference) [9]. For further details, let us firstly consider to have a fixed orientation and size of instudy object as shown in Fig. 2.
Here, we pick a so-called reference point (xc, yc), where:
which represent a straight line passing through (x, y). The couple (?, ?) is unique if ? ?[0, ?] and ? ? R or if ? ? [0, 2?] and ? ≥ 0. Each pixel in the original image is transformed in a sinusoid in the (?, ?) domain. The presence of a line is detected by the location in the (?, ?) plane where more sinusoids intersect.
Implementation of the SHT: SHT algorithm uses an array called accumulator to detect the existence of a line y = mx + b. The dimension of the accumulator is equal to the number of unknown parameters of SHT problem. For example, the Hough linear transform problem has two unknown parameters: m and b. The two dimension of the accumulator array would correspond to quantized values for m and b. For each pixel and its neighborhood, SHT algorithm determines if there is enough evidence of an edge at that pixel. If so, it will calculate the parameters of that line, and then look for the accumulator’s bin that the parameters fall into, and increase the value of that bin. By finding the bins with the highest value, the most likely lines can be extracted, and their (approximate) geometric definitions read off. The simplest way of finding these peaks is by applying some form of threshold, but different techniques may yield better results in different circumstances determining which lines are found as well as how many. Since the lines returned do not contain any length information, it is often next necessary to find which parts of the image match up with which lines. For more details, please refer to Ref. [6].
function off ?: in other words, we can build the R-table. The R-table allows us to use the contour edge points and gradient angle to recompute the location of the reference point. Let us remark that we need to build a separate R-table for each different object. Summarizing, after a quantization of the image space P[xcmin, ..., xcmax][ycmin, ..., ycmax], for each edge point (x, y) and using the gradient angle ?, we retrieve from the R-table all the (?, r) values indexed under ?. Then, for each (?, r), we compute the candidate reference point according to Eq. (4).
In the proposed procedure the RGB format has been used. Now, we increase a suitable counter measuring the votes for the considered reference point. Thus, possible locations of the object contour are given by local maxima in P[xc][yc] [10].
For further details about generalization of GHT in case of unfixed size or orientation, please refer to the proposed scientific bibliography.
makes possible to delete the unnecessary information, e.g. buildings, cars, human beings or animals, etc., and consider only the object of interest, i.e. the signs. Subsequently, the usage of the GHT allowed us to identify the specific road sign within its macro-class, by a comparison between the sign detect into the raw image and a collection of templates stored into a previously created“evaluation archive” (see Fig. 11). The whole procedure has been implemented within the Matlab c environment, and so it can be easily compiled as a DLL which can be used in an external GIS tool.
The performances of proposed approach are very encouraging: 97% of used road signs have been correctly recognized. Detailed results are shown in Tables 1 and 2. Here, it is possible to denote how a no parking sign has been incorrectly classified as a no stopping sign. Actually we are engaged in optimizing the procedure and linking the obtained DLL with a suitable GIS, in order to evaluate the proposed approach in a real-world application.
The production of GIS is oriented for the management of the road cadastre, the quality and the benefits of the services furnished, which do not depend only on the ability of the software component to extrapolate and make explicit the information contained in its databases in implicit form, but above all in the updating the same data bank, especially within the cases characterized by a specific dynamism.
In fact, the elements of the territory which must be represented (the horizontal and vertical system of traffic signs as well as the advertising posters) are characterized by a remarkable variety, involving both their spatial location and the alphanumeric attributes useful in order to their specialization. On the other hand, the integration of the system with a GIS concerning other territorial aspects is made possible only starting from a continuous and constant updating of the data bank related to the road cadastre, so avoiding the incongruence of the results from such an integration [1]. Nowadays, different commercial solutions are available in order to speed up the surveying of interesting elements along the roads. In this way, it is possible to collect a cartographic dataset by exploiting suitable hardware and photogrammetric software tools. The dataset is able to show the records of interesting elements along the roads in a very quick way, if compared to the classical topographic methods [2]. Nevertheless, even if the acquisition step is completely automated, the post-processing procedure is still now affected by a massive intervention of technicians, with a resulting slowdown for the surveying and positioning. Naturally, it becomes a problem in such fields given the automatic and real-time work requisite of such applications, e.g. if it is necessary fast interventions on roadways with, at the same time, an accurate consideration of the available economic funds [3]. The proposed approach aims to further investigate
? The exportation of the acquired data towards a GIS, where the last corrections will be carried out in order to respect the topological constraints, according to the already existing cartography.
computational level the GHT roughly consists in making each examined pixel in an image to“project” a copy of the searched pattern at various angles and scales (usually, projection and comparison takes place starting from the center of a certain object, but it’s possible to start elsewhere under special conditions) and then keeping track of how many pixel matches for a given scale and angle occurred between the “projection” and the tested image. Therefore, the most general algorithm definition says to do exactly that: creating a special reference data structure (usually a binary image, in the form of a table, called R-Table) and essentially comparing its “boundary” or “contour” with groups of pixels (having a central or boundary pixel for reference) [9]. For further details, let us firstly consider to have a fixed orientation and size of instudy object as shown in Fig. 2.
Here, we pick a so-called reference point (xc, yc), where:
which represent a straight line passing through (x, y). The couple (?, ?) is unique if ? ?[0, ?] and ? ? R or if ? ? [0, 2?] and ? ≥ 0. Each pixel in the original image is transformed in a sinusoid in the (?, ?) domain. The presence of a line is detected by the location in the (?, ?) plane where more sinusoids intersect.
Implementation of the SHT: SHT algorithm uses an array called accumulator to detect the existence of a line y = mx + b. The dimension of the accumulator is equal to the number of unknown parameters of SHT problem. For example, the Hough linear transform problem has two unknown parameters: m and b. The two dimension of the accumulator array would correspond to quantized values for m and b. For each pixel and its neighborhood, SHT algorithm determines if there is enough evidence of an edge at that pixel. If so, it will calculate the parameters of that line, and then look for the accumulator’s bin that the parameters fall into, and increase the value of that bin. By finding the bins with the highest value, the most likely lines can be extracted, and their (approximate) geometric definitions read off. The simplest way of finding these peaks is by applying some form of threshold, but different techniques may yield better results in different circumstances determining which lines are found as well as how many. Since the lines returned do not contain any length information, it is often next necessary to find which parts of the image match up with which lines. For more details, please refer to Ref. [6].
function off ?: in other words, we can build the R-table. The R-table allows us to use the contour edge points and gradient angle to recompute the location of the reference point. Let us remark that we need to build a separate R-table for each different object. Summarizing, after a quantization of the image space P[xcmin, ..., xcmax][ycmin, ..., ycmax], for each edge point (x, y) and using the gradient angle ?, we retrieve from the R-table all the (?, r) values indexed under ?. Then, for each (?, r), we compute the candidate reference point according to Eq. (4).
In the proposed procedure the RGB format has been used. Now, we increase a suitable counter measuring the votes for the considered reference point. Thus, possible locations of the object contour are given by local maxima in P[xc][yc] [10].
For further details about generalization of GHT in case of unfixed size or orientation, please refer to the proposed scientific bibliography.
makes possible to delete the unnecessary information, e.g. buildings, cars, human beings or animals, etc., and consider only the object of interest, i.e. the signs. Subsequently, the usage of the GHT allowed us to identify the specific road sign within its macro-class, by a comparison between the sign detect into the raw image and a collection of templates stored into a previously created“evaluation archive” (see Fig. 11). The whole procedure has been implemented within the Matlab c environment, and so it can be easily compiled as a DLL which can be used in an external GIS tool.
The performances of proposed approach are very encouraging: 97% of used road signs have been correctly recognized. Detailed results are shown in Tables 1 and 2. Here, it is possible to denote how a no parking sign has been incorrectly classified as a no stopping sign. Actually we are engaged in optimizing the procedure and linking the obtained DLL with a suitable GIS, in order to evaluate the proposed approach in a real-world application.