| Pavement surface distress affect the comfort and safety of driving.In order to improve the frequency and efficiency of pavement distress detection,it is urgent to develop the fully automatic pavement distress detection method,realize large-scale pavement distress detection,timely maintenance and repair,and improve the service life of pavement.At present,the multi-functional road inspection vehicle is used for the annual inspection of pavement distresss,which has low detection frequency and high cost.It is unable to realize the periodic high-frequency detection of pavement distresss.It is difficult to find seriously damaged pavement distresss in time,miss the best maintenance time for pavement distress,and increase the maintenance cost.Moreover,the road detection vehicle is equipped with large sensor equipments,which can not collect the pavement condition information of multiple lanes at the same time.Therefore,a lightweight detection equipment is urgently needed to complete the automatic pavement conditioin detection and assessment.As a portable,flexible and mobile flight equipment,UAV(Unmanned Aerial Vehicle)is widely used in infrastructure detection.It is mounted with high-resolution industrial camera for structure health monitoring.The UAV collects pavement distress images at a certain altitude.The flight parameters of the UAV will affect the imaging quality,and then affect the accuracy of pavement distress recognition.However,there is a lack of research on flight parameters,but setting UAV flight altitude information through manual experience can not meet the recognition accuracy of pavement distress.And the current automatic recognition algorithm shows poor universality,so it is unable to recognize and analyze pavement distress under different pavement structures and conditions.This paper built a pavement distress detection platform based on UAV to collect pavement distress images.Based on deep learning method,the pavement surface distress identification model is established to classify,localize,segment,measure and count various pavement surface distress,and evaluate the pavement condition.Firstly,the flight altitude and speed of UAV are studied to ensure that pavement distress images have sufficient resolution for identification and prevent motion blur.The pavement distress acquired by UAV is small-scale,and the types of distress are unbalanced,which seriously affects the establishment of pavement distress dataset.The small-scale target enhancement method of copy-paste is adopted to improve the distribution balance and category diversity of pavement distress.Because the full-scale pavement distress images are not convenient for model training,it is necessary to preprocess the pavement distress image collected by UAV,prevent the distortion of pavement distress images during scaling by Padding processing,and cropped them to512 scale to establish the asphalt pavement distress dataset UAPD(UAV aspahlt pavement distress dataset).Then,aiming at the problems of unbalanced pavement distress distribution and low identification accuracy of pavement distress,the YOLOv3 multi-scale model was used to identify and localize the pavement transverse cracks,longitudinal cracks,alligator cracks,oblique cracks,pothole and patch.Study the distress localization accuracy of different anchor sizes combinations and the distress extraction ability of different feature extraction networks,carry out the combination design of pavement distress region level identification network model,and complete the classification and quantitative statistics of pavement distress.The results show that the average classification and localization accuracy of six pavement distress of YOLOv3 based on Dark Net53 backbone feature extraction network is 56.6%,which can meet the requirements of pavement distress recognition.After the classification and localization of pavement distress,it is also necessary to segment and extract the morphology of pavement distress,and then complete the evaluation of pavement distress.Aiming at the problems of discontinuous segmentation results of pavement cracks and poor generalization ability in multi-scene conditions,a multi-feature fusion semantic segmentation w-segnet model was proposed.Using the low-level localization information and high-level semantic classification information of pavement distress,the pixel-level segmentation of pavement crack,patch and repair was realized,and the pixel-level segmentation accuracy of pavement distress in sunny days,shaded areas and wet areas was improved.The results show that the pavement distress segmentation model w-segnet based on encoding-decoding structure is 87.52%MPA and 75.88% MIo U in UAPD,which is better than the comparison model,and realized high-precision pixel-level segmentation of pavement crack,patch and repair.Finally,in order to calculate the pavement distress rate and complete the evaluation of pavement conditions,it is necessary to quantify the geometric characteristics of pavement distress and complete the statistics of pavement distress information.In view of the shortcomings of the current distress detection research,the inaccurate distress extraction and the imperfect automatic evaluation specification,the external regular rectangle,the minimum external rectangle and convex hull polygon were used to fit the boundary of pavement alligator crack,so as to solve the problem of inconsistent boundary of alligator crack calculation.Using the polar coordinates of the alligator crack connected domain to calculate the pavement alligator crack fragmentation,which made up for the defect of automatic calculation of pavement alligator crack fragmentation,and verified the accuracy of image-based pavement distress statistics by comparing with manual measurement.Results the average relative errors of pavement distress width,length and area were-0.931%,-1.160% and 0.573%,which met the requirements of engineering practice.To conclude,the research results of this paper can improve the fully-automated detection level of pavement distress,realize multi-dimensional and high-frequency lightweight pavement distress detection,promote the construction of smart highway,promote the application of advanced information technology,and gradually improve the full-element and full-cycle digital level of highway infrastructure maintenance and operation management. |
PDF Full Text Request