Timely and accurately detection as well as rehabilitation of road surface defects are of utmost importance for ensuring road safety and minimizing maintenance cost. However, the variety of pavement distress types and forms makes it difficult to accurately classify and detect them. To tackle the issue, this paper proposes a novel target detection model YOLO-SST based on YOLOv5 with the improvement in pavement distress features. First, a Shuffle Attention mechanism is introduced in the feature extraction backbone network to enhance the detection ability without significantly increasing the computational cost. Secondly, we add a detection layer and embed Swin-Transformer encoder blocks into the C3 module to capture global and contextual information. Finally, to improve the model’s detection ability, transfer learning is employed on a self-made dataset called RDDdect_2023, which consists of street view images captured via a DJI Action camera mounted on the car. Experimental results demonstrate that the YOLO-SST model outperforms YOLOv5 and other target detection models in terms of accuracy, recall rate, and mAP@0.5 value for detecting pavement distresses. This confirms that the YOLO-SST model has stronger feature extraction and fusion capabilities, resulting in better detection performance.

Timely detection and statistical analysis of pavement distresses are essential for improving road maintenance efficiency. However, traditional methods for pavement defect detection face challenges such as inefficiency and high equipment costs. In response to these challenges, this paper proposes a pavement defect detection and statistical system based on street view videos. Initially, we introduce an enhanced algorithm named SN-YOLO (Slim-neck YOLO) designed to address the issue of low model detection accuracy in complex background environments meanwhile achieve model lightweighting. Specifically, the GSConv lightweight convolution module is employed to minimize the model size, while the VoVGSCSP and VoVGSCSP-cheap modules are incorporated to augment the original C2f module, thereby refining the model’s recognition capabilities in intricate backgrounds. Moreover, by incorporating Soft-NMS for post-processing optimization, the model’s robustness in detecting multi-scale defects is enhanced. Experimental results on the open-source dataset RDD2022 and a proprietary dataset demonstrate that the improved SN-YOLO algorithm surpasses current state-of-the-art methods. Furthermore, by leveraging the SN-YOLO algorithm and the Deep oc-sort tracking algorithm, we develop a deployable pavement distress detection and statistic system. In the application to real-world road street view video analysis, the system exhibits unparalleled accuracy and efficiency in defect detection and data compilation, presenting a robust solution for expedited, large-scale assessment of pavement conditions.

Most advanced methods for road surface defect detection may have the issue of large network structures, making them impractical for implementation on mobile embedded systems. In this study, A lightweight road surface defect real-time intelligent detection system based on cloud-edge collaboration is proposed. The system comprises a lightweight feature-enhanced YOLO network (YOLO-LFE), a multi-object tracking network called ByteTrack, and is deployed on intelligent edge devices, enabling real-time defect detection and storage of uploaded defect information through edge computing modules and cloud systems. Built upon the latest YOLOv8 architecture, a lightweight MobileNetV3 network is first introduced as the backbone feature extraction component, resulting in a noteworthy decrease in parameters and computational intricacy. Subsequently, the Enhanced Spatial Pyramid Pooling (ESPP) method built upon the human visual perception system is applied to swap out the Spatial Pyramid Pooling Fusion (SPPF) method, thereby enhancing the model’s capability to capture features from tiny entities. Finally, an enhanced progressive feature fusion network is designed to further address the semantic gap issue during subsequent fusion of primary features. Experimental results on a self-made dataset and the public RDD2022 dataset show that the proposed detection model maintains high accuracy while requiring fewer parameters and less computational resources. Compared to YOLOv8, the model reduces the number of parameters by 32.5% and decreases computational requirements by 37%. This makes it more suitable for deployment on embedded devices. Additionally, the system has been field-tested on vehicles, thereby validating the effectiveness of the entire system.