VOL. 8, NO. 9, SEPTEMBER 2013 ISSN 1819-6608 ARPN Journal of Engineering and Applied Sciences © 2006-2013 Asian Research Publishing Network (ARPN). All rights reserved. www.arpnjournals.com 763 A NEW METHODOLOGY TO ESTIMATE DEFORMATION OF LONGITUDINAL SAFETY BARRIERS Marco Guerrieri 1 , Giuseppe Parla 2 and Ferdinando Corriere 2 1 Faculty of Engineering and Architecture, University of “Enna” Kore, Italy 2 Faculty of Engineering, University of Palermo, Italy E-Mail: marco.guerrieri@tin.it ABSTRACT This paper presents a new high-efficiency methodology to estimate deformations of longitudinal road safety barriers, whose monitoring is a necessary condition for the maintenance of high safety standards. The methodology is based on the analysis of videotaped sequences, obtained by means of two matched video cameras (“stereo head”) installed on a vehicle adequately equipped. Two different theoretical approaches have been defined: monoscopic and stereoscopic. The new methodology has been used to empirically evaluate the longitudinal safety barrier deformations on an Italian road. Keywords: longitudinal barrier, deformations, stereoscopic, image processing. DETECTION OF ROAD BARRIERS The first stage of procedure to estimate deformation of longitudinal safety barriers has been devoted to detect the extrados line in safety barriers; the Canny algorithm [1, 2] has then been used to reduce the image noise (edge detection). The Canny algorithm studies in detail the behavior of the gradient operator applied to a noisy edge. As a matter of fact this algorithm presupposes that the image edge front to be segmented is already corrupted by Gaussian white noise; therefore, such an algorithm convolves the image to be processed with a suitable Gaussian smoothing filter so as to meet the following conditions [1, 3, 4, 5, 6]:  high probability of detecting the real edge;  the image points highlighted by the gradient operator must be as faithful as possible to the real edge of the object under examination;  uniqueness of the answer (exclusion of any multiple edges of the object). The algorithm can be divided into different steps:  image Gaussian smoothing, applied separately to the two directions (x, y), to reduce the noise;  pixel gradient calculation;  thresholding or non-maximum suppression in longitudinal direction to the edge which utilizes two different threshold values. Such a phase is useful to distinguish the most marked (significant) edge points from those less marked (weak);  thinning (significant edge selection through hysteresis) which allows to eliminate or not some weak edge points in relation to the direction and the intensity of the gradient of the points adjacent to the angle. The high quality of the results from the Canny method can be justified by the fact that the method utilizes two thresholds, one for detecting the sharpest edges, the other for detecting the weakest edges. The latter are, however, taken into consideration only if they appear to be connected to “sharp” edges. Therefore, in Canny algorithm, the traditional single- threshold-based needs to be replaced with a double- threshold-based approach called hysteresis thresholding; the latter is employed after applying the non-maximum suppression. In particular, if we determine the two thresholds t 1 and t 2 as t 1 >t 2 , the algorithm establishes that the pixel located at position (i,j) on the generic chromatic plane p of the image A (i, j, p) is a Canny edge pixel if the intensity value of the pixel examined - A(i, j, p) - appears to be superior to the highest threshold value t1; while the lowest value t 2 indicates the limit under which the pixel is not a Canny edge without any doubt. Moreover, the algorithm determines that the pixel set with values ranging between t 1 and t 2 are edged only if they are connected to those certainly edged, or those whose value is higher than t 1 . The above can be summed up into the following inequalities: ە  ۔  ۓ Aሺi,j,pሻ> t1 then A(i,j,p) are edged t2 ≥ Aሺi,j,pሻ ≥ t1 then A(i,j,p) are defined as rough edges Aሺi,j,pሻ> t2 then A(i,j,p) isn’t an edge (1) It has also been observed that on the same image, an edge detection applied separately to the three chromatic planes (RGB) [1] produces, as expected, nearly similar results. In addition, the edges in all the objects in the scene, provided from the previous segmentation, differ in a few pixels only if the segmented images on the three planes are separately elaborated or the threshold results are grouped in a single plane, or else if the elaboration is done on a grey-scale image obtained by appropriately converting a color image. For these reasons we preferred to operate separately on the three chromatic image planes and group the processing results into one Boolean piece of information.