2021

Eyoab Zegeye Teshale, Minnesota Department of Transportation
Thomas Calhoon, University of Minnesota
Eddie Johnson, Minnesota Department of Transportation
Shongtao Dai, Minnesota Department of Transportation

Pavement tenting, also referred to as crack-heaving, is a distress condition that primarily affects bituminous roads constructed in cold climates. This type of distress spreads over long stretches of roadways and can drastically impact the safety and comfort of drivers. The phenomenon occurs in freezing winter temperatures offering a limited and dire time window for testing. This paper discusses using an integrated multi-sensor non-destructive testing methodology to evaluate and characterize pavements affected by tenting. A survey van equipped with high-definition video and thermal cameras, LIDAR laser scanner, high-resolution accelerometer, and ground-penetrating radar (GPR) technologies was used to assess several roads suspected of tenting. The plurality of measuring devices and the data fusion and synchronization capabilities proved useful in revealing important pavement tenting characteristics that would have been otherwise overlooked. The analysis of the data led to the development of test parameters, derived from longitudinal profile measurements, that captured reasonably well the intensity and frequency of the tented cracks. The parameters were successfully employed to characterize the tested roads and to determine the extent of critically affected segments. The study also showed the potential of GPR measurements to investigate underneath moisture conditions contributing to the formation of the tented cracks. Finally, the findings and the tools developed in this study were discussed and compared to observations of local engineers with extensive experience and insight on the subject matter. The knowledge and recommendations gathered in this final effort were also synthesized and incorporated in the paper.

Shuwei Li (shuweil@foxmail.com), Southeast University
Xingyu Gu, Southeast University
Xiangrong Xu, Nanjing Municipal Public Engineering Qualitu Inspection Center Station
Dawei Xu, Nanjing Municipal Public Engineering Qualitu Inspection Center Station
Tianjie Zhang, Zhejiang Scientific Research Institute of Transport
Qiao Dong, Southeast University
Zhen Liu, Southeast University

Concealed cracks are the cracks originating below the surface of asphalt pavement. These cracks could accelerate pavement damage and cause poor driving quality. However, the detection of concealed cracks has been a challenging task due to the nonvisibility of the crack. This study proposes an effective method to automatically complete the recognization and location of concealed cracks based on 3D ground penetrating radar (GPR) and the deep learning (DL) model. The original images of concealed cracks in three dimensions (B-scan, C-scan, and D-scan) could be obtained using 3D GPR. Then four processing steps were used to create the dataset, including filtering, recognizing, capturing, and labeling. The dataset consisted of 303 GPR D-scan images and 1306 cracks in total. After that, the You Only Look Once (YOLO) model was first introduced as the DL models. Eight models with three different versions (YOLOv3, YOLOv4, and YOLOv5) were selected to be trained and compared. The results reveal that this combined method is feasible for the detection of concealed cracks in asphalt pavement. Compared with YOLOv3, YOLOv4 and YOLOv5 both achieve the obvious progress even in a small dataset. The FPS of YOLOv4-tiny reaches 10.16 only using a medium CPU and the best mAP of YOLOv5x is up to 94.39%. According to detection results, YOLOv4 models have better robustness than YOLOv5 models and could distinguish between concealed cracks and pseudo cracks accurately. In conclusion, YOLOv4 models accomplish a good performance in detection speed and object confidence.

Jinchao Guan, Chang'an University
Xu Yang (yang.xu@chd.edu.cn), Chang'an University
Ling Ding, Chang'an University
Jingwei Liu, Monash University
Xiaoyun Cheng, Chang'an University

Automatic pixel-level pavement crack detection is facing huge challenges due to the crack irregularities, illumination variations, and other influencing factors. In order to efficiently complete the crack segmentation tasks in a practical environment, an automatic pixel-level pavement crack detection framework integrating stereovision technology and deep learning is presented in this paper. Automatic image acquisition and 3D pavement surface reconstruction are achieved by the vehicle-mounted photography system incorporating a multi-view stereo imaging technology. Based on the digital pavement model, we propose the 3D point cloud processing algorithm to generate the multi-feature image dataset consisted of color images, depth images and overlapped images, providing a new perspective for deep learning. To alleviate the computational burden, a U-net-based deep learning architecture introducing the depthwise separable convolution is proposed to conduct pavement crack segmentation, which allows for the use of small training dataset. This method was tested and verified in the asphalt roads within the city of Xi’an, China. The results show that the 3D pavement model achieves millimeter-level accuracy. Meanwhile, there are slight spatial differences between the 3D models generated by parallel photography and oblique photography. Using different types of images, three crack segmentation models were trained, namely 2D segmentation model, 3D segmentation model and 2D+3D segmentation model. The 2D+3D segmentation model performs better than other models. The overall F1 score, Precision and Recall of the 2D+3D model on the validation dataset are 0.860, 0.857 and 0.866, respectively.

Jiayun Huang, University of California, Berkeley
Nicholas Fournier, University of California, Berkeley
Alexander Skabardonis, University of California, Berkeley

The current Highway Capacity Manual (HCM) employs a simple 5-point system to assess the quality of bikeway pavement as part of the comprehensive bicycle Level of Service (LOS) evaluation. Unfortunately, the ambiguous and rudimentary nature of the existing Pavement Quality Index (PQI) fails to offer an objective review of bikeways across different jurisdictions. In the following analysis, first is an assessment of the current Pavement Quality Index and Bicycle LOS in the HCM. To demonstrate the impact of the pavement quality rating and the importance of a more standardized evaluation method, a sensitivity analysis is performed. Following, we propose an improved PQI matrix based on a comprehensive literature review. The new matrix allows for a more holistic understanding of pavement quality in a three-category framework. The proposed methodology includes specifications for the functionality, structural integrity, and maintenance of bikeways. Within each category, objective thresholds are defined, such as for potholes, cracks, and maintenance routines, in order to minimize any potential subjectivity.

Nima Safaei, University of Iowa
Omar Smadi, Iowa State University
Babak Safaei, Michigan State University
Arezoo Masoud, University of Iowa

Cracks considerably reduce the life span of pavement surfaces. Currently, there is a need for the development of robust automated distress evaluation systems that comprise a low-cost crack detection method for performing fast and cost-effective roadway health monitoring practices. Most of the current methods are costly and have labor-intensive learning processes, so they are not suitable for small local-level projects with limited resources, or they are only used for specific pavement types. In this paper, a new method is proposed that uses an improved version of the weighted neighborhood pixels segmentation algorithm to detect cracks in 2-D pavement images. This method uses the Gaussian cumulative density function as the adaptive threshold to overcome the drawback of fixed thresholds in noisy environments. The proposed algorithm was tested on 300 images containing a wide range of noise representative of different noise conditions. This method proved to be time and cost-efficient as it took less than 3.15 seconds per 320 × 480 pixels image for a Xeon (R) 3.70 GHz CPU processor to determine the detection results. This makes the model a perfect choice for county-level pavement maintenance projects requiring cost-effective pavement crack detection systems. The validation results were promising for the detection of low to severe-level cracks (Accuracy = 97.3%, Precision = 79.21%, Recall= 89.18% and F1 score = 83.9%).

Charalambos Kyriakou (kyriakou.charalambos@ucy.ac.cy), University of Cyprus
Symeon Christodoulou, University of Cyprus
Loukas Dimitriou, University of Cyprus

The paper presents a data-driven framework and related field studies on the use of supervised machine learning and smartphone technology for the spatial condition-assessment mapping of roadway pavement surface anomalies. The study introduced herein explores the use of data, collected by sensors from smartphones and automobiles’ on-board diagnostic (OBD) devices while vehicles are in movement, for the detection of roadway anomalies. The research proposes a low-cost and automated method to obtain up-to-date information on roadway pavement surface anomalies, with the use of smartphone technology, artificial neural networks (ANN), robust regression analysis and supervised machine learning algorithms for multiclass problems. The technology for the suggested system is readily available and accurate and can be utilized in pavement management systems (PMS) and geographical information system (GIS) applications. Further, the proposed methodology has been field-tested exhibiting accuracy levels higher than 90% and it is currently expanded to include larger datasets and a bigger number of common roadway pavement surface defect types. The proposed system is of practical importance since it provides continuous information on roadway pavement surface conditions, which can be valuable for pavement management systems and public safety.

Georgene Geary, GGfGA Engineering
Yichang Tsai, Georgia Institute of Technology (Georgia Tech)

3D pavement data is increasing in use and availability and opens up new opportunities to evaluate variability in pavements.  The majority of information we currently have on existing pavements is the result of the Long Term Pavement Performance Program (LTPP).  While the program is comprehensive and the data is immense, the study sections are typically only 500 ft in length, which limits the ability to accurately gauge the variability of the distresses in a pavement over a longer length, especially cracking in Jointed Plain Concrete (JPC) slabs.  3D pavement data already collected by transportation agencies has the opportunity to complement LTPP data to analyze variability and improve the use of LTPP data.  This paper presents a unique method to complement LTPP data using 3D pavement data, consisting of four steps: 1) crack detection using 3D pavement data, 2) categorize detected cracks by orientation and extent by slab using 3DSBM (3D slab based methodology), 3) convert categorized slab level cracking into Pavement ME Design (PMED) cracking, and 4) perform Local Calibration with the 3D converted PMED input values.  The method utilizes 3D pavement data to provide a non-discrete value for percent cracking in GPS-3 LTPP sections for the purposes of local calibration.  The proposed method is shown to be feasible using 3D pavement data and two JPC LTPP sections in Georgia.  The method could be extended to any of the State DOTs that have active LTPP sections and are now or will shortly be collecting 3D pavement data.

Eyal Levenberg, DTU Byg
Asmus Skar, DTU Byg
Shahrzad Pour, Danmarks Tekniske Universitet
Ekkart Kindler, Danmarks Tekniske Universitet
Matteo Pettinari, Vejdirektoratet
Milena Bajic, Danmarks Tekniske Universitet
Tommy Alstrøm, Danmarks Tekniske Universitet
Uwe Schlotz, Sweco Danmark

Modern cars are equipped with a great number of sensors that measure information about the vehicle and its surrounding. Thus, many of these measurements are related to the ride-surface conditions over which the vehicle is passing. In the paper, a four-component vision was outlined for advancing the idea of performing road condition evaluation based on in-vehicle sensor readings, and the subsequent feeding of Pavement Management Systems (PMSs) with live condition information. So doing is expected to further augment and enrich the functionalities of PMSs, and ultimately lead to improved maintenance and repair decisions. Presented next was the LiRA project -- an ongoing proof-of-concept attempt to realize the four vision components. The project elements and software architecture were described in detail, listing any assumptions, compromises, and challenges. LiRA involves a fleet of 400 electric cars operating in Copenhagen, both within the city streets and nearby highways. Sensor data collection was performed with a customized Internet of Things (IoT) device installed in the cars. Data processing and structuring involved new software tools for quality control, spatio-temporal interpolation, and map matching. A hybrid approach, combining machine learning models with physical models, was applied to convert data into relevant information for PMSs. Based on the experience with LiRA, the vision actualization is deemed achievable, workable, and upscalable.

Radhika Ravi, Purdue University
Darcy Bullock, Purdue University
Ayman Habib, Purdue University

Regular pavement monitoring over highways and airport runways is vital for public agencies to ensure safety of vehicles and aircraft. Highways are subject to cracking and potholes along with some instances of debris around work zones. Airports are also concerned with debris but have much lower tolerance for the presence of foreign object debris (FOD) that could possibly damage aircraft. LiDAR is emerging in a variety of mobile mapping systems (MMS) and will likely be integrated into many transportation vehicles over the next decade for pavement inspection. This paper proposes a unique algorithm for surface inspection with the help of MMS driven at highway speeds. The study analyzed LiDAR data for 8 miles of highway collected at approximately 55mph, which indicates that an adequately designed MMS along with the proposed algorithm can detect pavement anomalies as small as 2cm including cracking, potholes, and/or surface debris. This is more than sufficient for highways, where debris such as ladders and tires are an order of magnitude larger. For evaluating the effectiveness of detecting smaller airport FOD, a validation dataset was created by driving the MMS at 15mph adjacent to a debris field of 50 pieces of FOD collected from an airport. The study found that 100% of the FOD items larger than 2cm in size (12 out of 50 samples) were detected successfully. Both datasets suggest that MMS LiDAR is sufficient for pavement inspection and as sensor fidelity increases, even small FOD can be detected with the algorithm proposed in this paper.

Rohit Ghosh, Iowa State University
Omar Smadi, Iowa State University

Pavement distresses lead to pavement deterioration and failure. Accurate identification and classification of distresses helps agencies evaluate the condition of their pavement infrastructure and assists in decision-making processes regarding pavement maintenance and rehabilitation. The state of the art is automated pavement distress detection using vision-based methods. In this study, we implemented two deep learning techniques, Faster R-CNN and YOLO v3, for automated distress detection and classification of high-resolution (1,800 × 1,200) 3D asphalt and concrete pavement images. Our training and validation dataset contained 625 images that included distresses manually annotated with bounding boxes representing the location and types of distresses and 798 no-distress images. Data augmentation was performed to enable more balanced representation of class labels and to prevent overfitting. YOLO and Faster R-CNN achieved 89.8% and 89.6% accuracy respectively. Precision-recall curves were used to determine the Average Precision (AP), which is the area under the precision-recall curve. The AP values for YOLO and Faster R-CNN were 90.2% and 89.2% respectively, indicating strong performance for both models. We also developed Receiver Operating Characteristic (ROC) curves to determine the area under the AUC curve, another indicator of model performance. AUC values of 0.96 for YOLO and 0.95 for Faster R-CNN also indicate our models perform well. Finally, we evaluated our models by developing confusion matrices comparing our results with manual QA/QC. Our models’ very high level of match, namely 97.6% match to manual QA/QC for YOLO and 96.9% for Faster R-CNN, suggest our methodology has potential as replacement for manual QA/QC.

Yung-An Hsieh, Georgia Institute of Technology (Georgia Tech)
Yichang Tsai, Georgia Institute of Technology (Georgia Tech)

Raveling is one of the most common asphalt pavement distresses. The survey of its condition is required for transportation agencies to not only ensure roadway safety but also appropriately apply preservation and rehabilitation treatments. However, the traditional raveling condition survey, including the determination of the raveling severity, is manually conducted through in-field visual inspection methods, which are time-consuming, labor-intensive, and error-prone. Although automated raveling detection and severity classification models have been developed, these existing models have their own shortcomings. Therefore, there is an urgent need to develop a more accurate and reliable model to automatically detect and classify raveling. In this study, we propose a convolutional neural network (CNN)-based model for automated raveling detection and classification. Compared to general CNNs, the proposed model combines the data-driven features learned from the training data with the macrotexture features of 3D pavement surface data to achieve better performance. The proposed model was evaluated and compared with the baselines using real-world 3D pavement surface images collected from the State of Georgia. By utilizing both macrotexture and data-driven features, the proposed model achieved the highest accuracy of 90.8%. The proposed model also achieved precisions and recalls higher than 85% in all raveling severity levels, which are more balanced and higher than the baselines. We concluded that the utilization of image features largely affects the performance of models on automated raveling detection and classification.

Pingzhou Yu, Georgia Institute of Technology (Georgia Tech)
Yichang Tsai, Georgia Institute of Technology (Georgia Tech)

Pavement raveling is one of the predominant distresses in the US that impact roadway safety and driver comfort on open-graded friction course (OGFC) pavements. It is difficult to reliably forecast the raveling deterioration on OGFC for applying the right treatment such as milling only the OGFC layer, with the current qualitative condition assessment method (Severity Level 1-3 or Light, Moderate, Severe). There is an urgent need to develop a method to quantitatively evaluate raveling condition. This paper proposes a method with loss of aggregate as a new performance indicator to automatically quantify raveling using 3D pavement surface data already collected by transportation agencies for pavement evaluation. The proposed method consists of 1) 3D data acquisition, 2) pre-processing with a) outlier removal and image smoothing, b) two-sensor image stitching, c) range image rectification, and 3) raveling detection using a) region of interest selection, b) reference surface estimation, c) potential aggregate loss identification, d) noise removal and; 4) aggregate loss quantification. The proposed method is validated using pavement images (with known aggregate loss) from simulated pavement mats fabricated in the lab, and synthetic pavement images obtained through procedural generation. Validation results show that a strong correlation (r=0.99) between the computed aggregate loss and ground references. A better performance is observed in the proposed method compares to other methods (watershed method and model fitting method). The proposed method provides a cost-effective mean to quantify loss of aggregates in support of quantitative raveling condition forecasting by leveraging the 3D pavement data already collected by transportation agencies.

Ye Ma, Louisiana State University
Mostafa Elseifi (elseifi@lsu.edu), Louisiana State University
Nirmal Dhakal, Louisiana State University
Mohammad Bashar, University of Colorado, Boulder
Zhongjie Zhang, Louisiana Department of Transportation and Development

Ground Penetrating Radar (GPR) is a nondestructive evaluation (NDE) technique, which has been applied to assess as-built pavement conditions and to evaluate damage and deterioration that develop over time.  The objective of this study was to develop a methodology that uses GPR to detect moisture-related stripping damage in asphalt pavements. To achieve this objective, A Finite-Difference Time-Domain (FDTD) based simulation program was used to study the propagation of GPR signals in a stripped pavement. Field test data including GPR scans and visual inspection of cores of 204 pavement sections were used to study the relationship between GPR traces and AC stripping damage. Based on this analysis, a novel GPR-based indicator, known as the Accumulating In-layer Peaks (AIP), was introduced to detect stripping damage in asphalt pavements.  Field data and pavement cores were used to validate the proposed indicator and to evaluate its effectiveness in detecting the presence, extent, and severity of stripping in in-service pavement sections.  Based on the results of the study, it was found that the presence of a void in middle of the AC layer resulted in positive peaks in the reflected waves as indicated by the simulation of GPR signals. In addition, detected intermediate wave peaks between the surface and the interface between the AC and base layers on the GPR traces were associated with stripping damage in the AC layer. The AIP predicted accuracies for stripped and non-stripped sections were 80% and 96%, respectively, indicating its effectiveness to detect stripping damage in flexible pavements.

Danilo Balzarini, International Cybernetics Co.
James Erskine, International Cybernetics Co.
Michael Nieminen, International Cybernetics Co.

The development of new laser technologies in recent years has changed pavement data collection, opening the door to a fully automated approach. In this paper the application of the Pavement Surface Cracking Metric (PSCM), inspired by the Universal Cracking Indicator proposed by William Paterson in 1994, and developed by the ASTM E17 group is presented. The method uses quantitative definitions to ensure consistency of the results and eliminate the subjectivity associated with human rating of the pavement distresses. Multiple runs of pavement data have been collected on three asphalt sections to assess the repeatability and reproducibility of the method. The Pavement Surface Cracking Index is also introduced to convert the PSCM value, which is a physical property of the pavement, into a 100-0 score of the pavement section. Finally, the use of the PSCM to classify the pavement distress and the inclusion of potholes and patching in the metrics are discussed.

Jiawei Gao, Northern Arizona University
Chun-Hsing Ho, Northern Arizona University
Igor Wiese, Northern Arizona University
Dada Zhang, Northern Arizona University
Marco Gerosa, Northern Arizona University

A novel approach using machine learning based technology is presented for pavement condition assessment and prediction. A year long vibration data collected in the I-10 corridors located in Pheonix, Arizona USA was obtained for analysis. All vibration data were analyzed through three steps: cluster analysis, resampling, implementing and evaluating machine learning algorithms. Cluster analysis eliminates high correlation parameters to improve algorithm efficiency. Resampling avoids the possible over-fitting results of an unbalanced dataset. We implement and evaluate commonly used machine learning algorithms, including Logistic Regression (LR), Support Vector Machine (SVM), Random Forest (RF), and Neural Network (NN), for the evaluation of three pavement conditions (i.e., slight, moderate, and severe) and future predictions. Among the four computing algorithms, it is found that Random Forest achieved the best performance for condition assessments and performance predictions of highway pavements with its rated accuracy of 98%, a Matthews correlation coefficient of 76.54%, a precision of 95%, and a recall of 77%. The results show that machine learning algorithms based on random forest can provide accurate pavement condition detection. The methodology presented in the paper could help highway agencies and research insititions to determine the levels of pavement deterioration and develop road maintenance plans.

Afshin Famili, Texas Department of Transportation
Wayne Sarasua, Clemson University
Alireza Shams, SUNY Farmingdale
William Davis, Citadel Military College
Jennifer Ogle, Clemson University

Periodic measurement and identification of the presence and severity of pavement rutting is crucial for pavement management programs conducted by state transportation agencies. This paper proposes a novel analytical method for identifying pavement rutting locations using data collected by mobile terrestrial LiDAR scanning (MTLS).  Four vendor MTLS systems are evaluated based on their ability to accurately reproduce a roadway’s transverse profile.  For the purpose of establishing ground truth measurements, 2-inch interval pavement transverse profiles, which included rutting sections, were collected using traditional surveying techniques. MTLS transverse profiles were evaluated using partial curve mapping (PCM), Frechet distance, area, curve length, and Dynamic Time Warping (DTW) techniques. Resultant pavement transverse profiles were compared between vendors and a profile created from traditional surveying.  Results indicate calibrated MTLS systems can provide accurate transverse profiles for potential identification of pavement rut areas.  Based on this determination, a novel method was developed for use in identifying locations of pavement rutting through analysis of the curvature of MTLS raster surfaces. After evaluating three grid cell sizes for elevation raster surfaces, a 1-foot by 1-foot raster grid cell size was determined to be most suitable for identifying continuous concave raster cell groups along wheel path trajectories.  These cell groupings were found to reliably identify pavement rutting locations.  The analytical procedures employed through application of this method comprise an efficient work flow process that is not reliant upon a time-consuming continuous comparison with a MTLS modeled uniform surface. Keywords: Mobile terrestrial LiDAR scanning, transverse profile, pavement rutting identification