ABSTRACT Deep Learning (DL) techniques have been applied to the processing and evaluation of pavement conditions. DL-based predictive models are rapid and have higher accuracy compared to traditional methods. In this research, some models were developed to appraise pavement patching by considering the severity and the manhole cover utilising images collected by the Road Surface Profiler (RSP) from Bus Rapid Transit (BRT) lines in Tehran, Iran. Two cases (scenarios) were regarded in this research: 1) Patching without considering the severity, and 2) Patching by considering the severity. YOLOv5, YOLOv6, and YOLOv7 as one-stage object detection algorithms, and Faster R-CNN (ResNet-50) and Faster R-CNN (MobileNet_v3) as two-stage object detection algorithms were evaluated to choose the desired algorithm. The results exhibited that the trained model on YOLOv5 has better performance and is faster than the rest of the algorithms. In addition, the influence of the classes’ combination on model performance was scrutinised. Based on the results of the other classes, the outcomes of the study suggested that the model can detect and train better when the model is trained by a dataset with all classes. Furthermore, the performance of the model was also investigated by examining the effect of image size and YOLOv5's architectures. Based on the obtained results, YOLOv5 and an image size of 640*640 outperformed the rest. In Case 1, the precision, recall, F1-score, and mAP were 75.8%, 67.4%, 71.35%, and 77.4%, respectively. Moreover, the values of these metrics in Case 2, were 70%, 61.2%, 65.3%, and 64.4%, respectively. Hence, the results demonstrated that the developed model can be used with high performance in the detection and evaluation of pavement patching by severity, and manholes in Pavement Management Systems (PMS) and maintenance and rehabilitation plans.
Read full abstract