Road surface pothole detection is crucial for ensuring the driving safety and path planning of autonomous vehicles. However, existing detect
Road surface pothole detection is crucial for ensuring the driving safety and path planning of autonomous vehicles. However, existing detection methods are often affected by variations in lighting, weather conditions, and complex environments, resulting in lower detection precision and recall rates. To address this, this paper proposes an innovative improved algorithm, which is based on the YOLOv8 model, and introduces the MSF-HFEB module in the innovative design. By cleverly blending the strengths of Convolutional Neural Network (CNN) and the Transformer architecture, the algorithm accomplishes multi-scale feature extraction and fusion for both local and global aspects of road potholes. This design significantly improves the robustness of the algorithm under different lighting and complex environmental conditions. Additionally, the algorithm incorporates the LSKA attention mechanism and an integrated MHSA_CGLU module. By utilizing large-scale convolutional kernels for weighted feature screening, it further enhances the capability of multi-scale feature extraction and the richness of nonlinear feature representation. Experiments on public datasets have validated the effectiveness of our approach: compared to the original YOLOv8 model, the detection precision and recall rate of our algorithm have been improved by 9.8% and 11.6% respectively, and the F1 score has reached 84.9%. The outcomes of this research not only enhance the driving safety and comfort of autonomous vehicles in complex environments but also offer robust technical support for obstacle avoidance and path planning in self-driving vehicles.
