Detecting Pavement Distresses Using Crowdsourced Dashcam Camera Images

University Transportation Centers Program2021PDFTech ReportDadashova, BaharDobrovolny, Chiara SilvestriTabesh, MahmoodSafety through Disruption (Safe-D) University Transportation Center (UTC)Texas A&M Transportation InstituteUnited States. Department of Transportation. University Transportation Centers (UTC) ProgramUnited States. Department of Transportation. Office of the Assistant Secretary for Research and TechnologySafety through Disruption (Safe-D) University Transportation Center (UTC)United StatesCrowdsourcingDetection and identification systemsImage analysisMachine learningPavement crackingMachine visionVideoPavement Damage DetectionObject DetectionDashcamTTITTI- Student \u2013 0769A3551747115/TTI-Student-07Final Research ReportPavements play a vital role in the transportation infrastructure in the United States. Damage to public road transportation infrastructure causes roadways to fail to perform as intended and increases crash risks. Road damage must be detected quickly and accurately in order to maintain roads and effectively allocate repair money. In this study, four deep-learning object detection models were used for detecting five types of pavement damages using Nexar Dashcam images. The single-shot multi-box detector (SSD) and faster region-based convolutional neural networks (Faster R-CNN) object detection models using MobileNet and Inception techniques were applied to the selected and processed images. An attempt was made to detect alligator cracking, longitudinal cracking, transverse cracking, and patching on the roadway surface using the developed model. A total number of 6,326 images with a total number of 8,970 damage instances were selected from the provided data. Faster R-CNN models showed higher recall and precision compared to the SSD models. Faster R-CNN Inception Resnet showed the highest average precision of 56%, followed by Faster R-CNN Inception with an average precision of 50%. Recall values showed almost the same trend as precision. Faster R-CNN showed the highest average recall of 43%. The results showed that Type-I and Type-II damages were detected with the highest recall and precision compared to other damages. Type-I damage was detected with an average precision of 81% and recall of 70% in all of the models.127