| In recent years, with the expanding of China’s road network, road maintenance andmanagement issues have become increasingly prominent. In order to meet the requirements oflarge-scale, high efficiency and quality of road maintenance management, PavementManagement System (PMS) has been widelypromoted,which changed the traditional roadmanagement mode and made highway management decision-making more objective andscientific. The effectiveness of the pavement management system is dependent on accuratedata, and the pavement distress data is one of the key indicators to evaluate the status ofroad quality. At present, domestic and foreign research and developed a variety of pavementimage acquisition and processing systems, these systems can capture the pavement imageson the line, and process the images off the line.There problems such as uneven illumination,poor versatility and efficiency of the identification algorithm have become the" bottleneck"affecting the promotion and application of the system.During the past40years, the automatic pavement distress image identificationtechnology has been a classic problem on the field of traffic information engineering andpattern recognition. Because the pavement images have complex textures, object types, objectsignal and illumination, it is harder to identify them than other images.Existing algorithms aredeveloped on the pavement image with good qualityand clear cracks, which is not suitable tocomplex environments, it is difficult to meet the needs of engineering applications.Aiming at the problems above, on the basis of both domestic and international research,this paper proposed a new automatic pavement distress identification algorithm based onmulti-feature fusion technology. This paper selected the pavement distress as the researchobjects such as road marks, road cracks, loosen broken and surface patch. Through analyzingof the grayscale, texture, edges and geometry shape of the distress objects, severalfeature-fusion models are established in this paper. With these models, good segmentationresults have been achieved.This research has fruits and innovation as follows:(1) This desertation proposed a new pavement distress image processing flow of“classification before segmentation", its processing order is contrary to that from thetraditional methods. The proposed flow employs low complexity algorithm for qualitativeanalyses of the images. After the analyses, the images are classified into4categories: theintact pavement, loosen broken, cracks and patches. Then the high complexity algorithms areemployed for quantitative analyses of the different images. Finally, the post processes such asrecognition, segmentation, measurement and evaluation of the pavement distresses are done.This process flow can greatly improve the efficiency and accuracy of the algorithm for theidentification of the pavement distress objects, which makes various algorithm modules havebetter pertinence.(2) A new extraction algorithm fused the edge and grayscale characteristics of the roadmark images is proposed, the algorithm uses beamlet transform to extract the straight edge ofa binary image, which realizes the rapid screening of the pavement images including road marks. With the extracted straight edges, a pavement image can be segmented for manyregions. Finally, according to the grayscale characteristics of these regions, the split-mergealgorithm is used to segment the road mark area accurately. In experiments, the results showthat the efficiency and detection accuracy of the proposed algorithm are much higher than thatof the dynamic threshold segmentation algorithm.(3) A new shape descriptor–the Equivalent Length of the Connected Component isproposed. This descriptor gets the abstract outline of shape objects based on the Fourierdescriptor reconstruction method, and then the medial axis transform is used to gain themedial axis of the shape. Finally, the length of the medial axis is corrected by the object’sroundness. Experimental results show that this descriptor is very suited to distinguish thelinear target from the convex polygon.(4) A pavement distress initial classification algorithm is proposed based on the fusion oftexture and shape features. First, the proposed algorithm fuses the local contrast enhancedimagePt with the global grayscale corrected imageI ’tto get an enhanced pavement distressimage. Secondly, the enhanced image is decomposed with three-layer wavelet transform,which obtains the texture features of the whole image. Thirdly, an improved P-tile method isused to obtain the binarization image. From the binarization image, three shape features areextracted such as AA (the Average Area of all Connected Components), AM (the Area of theMaximum Connected Component), EL (the Equivalent Length of the longest ConnectedComponent). Eventually the neural network is used to fuse the texture and shape featuresserially. This algorithm achieves an effective classification of the pavement distress image,and its accuracy is higher than the algorithm with single type features.(5) A crack target detection algorithm fused grayscale, ridge edge and shapecharacteristic is proposed. First, a pavement crack image is divided into manynon-overlapping sub-blocks, and then a histogram estimation method is used to take theoptimal threshold of each sub-block. Through these thresholds,the grayscale feature image ofthe pavement cracks is obtained. Secondly, the edge feature image of the pavement cracks isgot based on multi-scale ridge edges fusion algorithm, and then pixel-level fusion is operatedon the feature images with OR operation. Thirdly, the noise of the pixel-level fused image isremoved through the DS evidence theory and shape analysis method. At last, the improvedconnection algorithm is adopted to achieve the accurate segmentation of the crack target. Theexperimental results show that the segmentation results of the proposed algorithm are superiorto that from the other three classic segmentation algorithms.All the algorithms and methods proposed in this desertation have been realized in thereal developed system. |
PDF Full Text Request