Accurate and efficient inspection of road surface defects is essential for ensuring traffic safety and supporting timely maintenance. However, existing vision-based inspection techniques often rely on high-resolution cameras or laser sensors, which limit robustness under complex or dynamic environmental conditions. To overcome these challenges, this study develops a vehicle-mounted acoustic inspection framework based on tire–road coupling noise for rapid road surface condition assessment. An acoustic inspection dataset is established, covering multiple types of urban pavement defects under diverse traffic and environmental conditions. The captured noise signals are transformed into MFCC, which serve as frequency-domain features for defect detection. A deep neural network integrating CNN, SE attention, and BiLSTM modules is developed to extract multi-scale time-frequency features and model temporal dependencies from tire-road coupling noise for pavement defect detection. Model evaluation results demonstrate that the proposed CNN-SE-BiLSTM model can accurately classify pavement defects without reliance on visual sensors, achieving an overall F1-score of 84.0%, while maintaining real-time inference and strong robustness to varying road surface conditions. Compared with existing vision- and Sensor- based inspection methods, the proposed method offers advantages in lower deployment cost, reduced sensitivity to illumination and weather variations, and easier integration into standard vehicles for continuous large-scale inspection. A field experiment on urban roads verifies the effectiveness of the proposed road surface inspection method, and a total of 12 pavement defects, including 9 cracks and 3 potholes, were successfully identified under real driving conditions. It is concluded that the proposed tire-road coupling noise method provides a cost-effective solution for road surface inspection. ... Read more

Accurate segmentation and quantitative characterization of pavement cracks are critical for road condition assessment and preventive maintenance. However, existing methods often lack robustness under complex field conditions, leading to inaccurate estimation of crack length and width. In this study, a two-step method is proposed for automatic crack segmentation and characterization using a vehicle-mounted system. Firstly, a high-resolution pavement dataset comprising 10,348 images, augmented with diverse environmental conditions, is established for model training and evaluation. Secondly, an improved SegFormer network with coordinate attention is trained and employed to enhance crack boundary preservation and suppress background noise in segmentation. Thirdly, an improved A* algorithm integrated with a dynamic window approach (DWA) is applied to extract continuous crack centerlines and adaptively compute length and width through perpendicular distance measurements. Experimental results demonstrate that the proposed method achieves superior performance with an accuracy of 98.74%, mPA of 85.79%, and inference speed of 149 frames per second, outperforming traditional segmentation model. Field validation further confirms that the relative error of crack length and width estimation is lower than 10%. These findings indicate that the proposed two-step method provides an accurate, efficient, and robust solution for real-time pavement crack characterization in practical road inspection scenarios. ... Read more

This paper proposes a nondestructive evaluation method based on deep learning using combined ground-penetrating radar (GPR) and electromagnetic induction (EMI) data for autonomic and accurate estimation of the cover thickness and diameter of reinforcement bars. A real-time object detection algorithm—You Only Look Once–version 3 (YOLO v3)—is adopted to automatically identify the reinforcement bar reflected signals from radargrams, with which the range of the cover thickness is roughly predicted. Subsequently, EMI data, accompanied with the cover thickness range, are imported to a one-dimensional convolutional neural network (1D CNN), pretrained by calibrated EMI and GPR data, to simultaneously estimate the cover thickness and reinforcement bar diameter. Testing with the on-site GPR data shows that YOLO v3 is superior to Single Shot Multibox Detector method in GPR hyperbolic signal identification. Testing of 1D CNN with the EMI and GPR data collected in an in-house sand pit experiment shows that the estimation accuracy of the cover thickness and reinforcement bar diameter is, respectively, 96.8% and 90.3% with a permissible error of 1 mm. Further, an experiment with concrete specimens demonstrates that among the 22 estimated values (including the reinforcement bar diameter and cover thickness), there are 17 values accurately estimated, while the inaccurately estimated values have an error up to 2 mm. The experimental results show that the proposed method can autonomically evaluate the reinforcement bar diameter and cover thickness with a high accuracy. ... Read more