Involution and Kolmogorov-Arnold networks based computationally efficient algorithm for detection of road defect with use of oriented bounding boxes

Effective management of road infrastructure is a cornerstone of sustainable transport development and road safety. One of the key challenges in this field is developing accurate and reliable methods for automatic detection of pavement damage that can operate effectively under diverse environmental conditions and with limited computational resources. This study aims to develop an efficient embedded system for real-time road pavement condition monitoring, designed for integration at the vehicle level. To achieve this, a novel neural network model—INvolution Kolmogorov-Arnold based Detection (INKA-Det)—was proposed, featuring an innovative architecture that combines involution mechanisms with the Kolmogorov-Arnold representation theory to deliver high detection accuracy while maintaining low computational cost. A distinguishing feature of the proposed solution is the use of Oriented Bounding Boxes (OBB), which enable precise alignment with the orientation and shape of pavement defects, significantly improving the detection of longitudinal and transverse cracks. Additionally, a new open dataset with OBB-based annotations was created to support the training and evaluation of detection models. The study also introduces a dedicated hardware architecture for deploying the system in vehicles, enabling real-time data acquisition and on-device inference. Experimental results confirmed the robustness and effectiveness of the proposed approach — the INKA-Det model outperformed state-of-the-art solutions, achieving higher detection accuracy than model from You Look Only Once (YOLO) family, while preserving computational efficiency. The developed system represents a significant step toward practical, autonomous, and scalable road pavement condition monitoring.