With the aging of global transportation infrastructure and the increasing demand for intelligent maintenance, real-time monitoring of road surface defects faces three core challenges: large scale differences, strong background interference, and difficulty in edge deployment. Traditional manual detection is inefficient and highly subjective, and existing deep learning models struggle to balance accuracy and real-time performance. This research proposes a real-time monitoring system based on an improved lightweight YOLOv8, achieving breakthroughs through three innovations: dynamic feature pyramid, spatial-channel dual-dimensional attention, and dynamic pruning. The system combines the TensorRT engine with a heterogeneous computing architecture, achieving 54 FPS@720p real-time inference on the NVIDIA Jetson platform, with end-to-end latency < 33 ms and stable power consumption of 8.3 watts. Validation was performed using a multi-source dataset containing 18,000 images: 4,227 labeled images (integrating the RDD2020 public dataset and five provincial self-built tunnel defect datasets, divided into training/validation/test sets in a 6:2:2 ratio), and 13,773 unlabeled images (used for semi-supervised pre-training and system-level real-world scenario validation). The results show that the model achieves an mAP of 78.5%, which is 3.2 percentage points higher than YOLOv8n. The rut depth error is 6.2 mm, and the robustness in extreme scenarios (low light in tunnels, reflective rain) is significantly better than traditional methods. This study provides a high-precision, low-power solution for the digital transformation of road maintenance. The innovation lies in the algorithm-hardware co-optimization framework and multimodal adaptive design.
Hao et al. (Fri,) studied this question.