Orthogonal and geometric feature learning for pavement damage detection in multi-platform imagery

Automated pavement damage detection is essential for infrastructure maintenance, but existing methods treat all damages as generic objects, neglecting the directional characteristics of certain defect types such as longitudinal and transverse cracks, which limits detection accuracy. This study presents an integrated framework combining orthogonal and geometric feature learning for pavement damage detection using multi-platform imagery. The framework incorporates three key components: an orthogonal convolution module that captures horizontal and vertical directional features through parallel branches, a geometric IoU loss function with aspect-ratio-driven dynamic weighting for improved localization of elongated damages, and an adaptive shape-based classification mechanism leveraging geometric properties to distinguish damage types by orientation. Evaluated on the RDD2022 dataset containing UAV aerial imagery and vehicle-mounted ground cameras from multiple countries, our method achieves mAP50 of 0.612 and F1 score of 0.613, with precision of 0.674, demonstrating superior performance in detecting orientation-sensitive damages. Cross-platform validation confirms robust generalization across different imaging systems and geographical regions, making the framework suitable for practical deployment in pavement health monitoring and infrastructure asset management. • Orthogonal convolution module captures horizontal and vertical pavement damage features. • Geometric IoU loss with aspect ratio weighting improves elongated crack detection. • Achieves mAP50 of 0.612 on RDD2022 with superior directional damage performance. • Adaptive classification leverages geometric properties to distinguish crack types. • Cross-platform validation demonstrates robust generalization across UAV and vehicle cameras.