SDSNet: A Multi-Task Learning Framework for Simultaneous Detection and Segmentation in UAV-Based Pavement Inspection

UAV-based pavement inspection is of great significance for advancing low-altitude intelligent transportation systems. However, most existing methods focus on only a single task, such as pavement distress detection, road area segmentation, or lane line segmentation, without addressing them within a unified framework. To address this limitation, this paper proposes SDSNet, a multi-task learning framework for simultaneous detection and segmentation in UAV-based pavement inspection. SDSNet achieves comprehensive pavement condition monitoring by simultaneously performing pavement distress detection, road area segmentation, and lane line segmentation within a unified framework. Specifically, a scale-adaptive knowledge aggregation module is designed to integrate task-specific features, providing more discriminative high-level features for multi-task learning. Moreover, a selective gated fusion module is introduced to adaptively fuse heterogeneous task features, facilitating efficient cross-task information interaction and improving feature complementarity. Finally, a multi-task PAN is constructed by extending the pyramid aggregation network to multi-scale representations that accommodate the distinct requirements of each sub-task. Experimental results demonstrate that SDSNet achieves competitive performance across multiple tasks, highlighting its effectiveness and practical potential for UAV-based automated pavement inspection. Note to Practitioners—UAV-based pavement inspection offers a flexible and efficient approach to road condition monitoring. Its advantages in spatial coverage and road accessibility make it complementary to established vehicle-based inspection methods. However, in practical deployments, UAV inspection systems often rely on separate algorithms for different tasks, such as pavement distress detection, road area segmentation, and lane line segmentation. This fragmented design leads to redundant computation, inconsistent outputs, and difficulties in real-time onboard processing. This paper presents a unified solution that allows a single model to perform multiple inspection tasks simultaneously. Through cross-task information sharing, the proposed method improves the consistency and comprehensiveness of pavement perception. The framework is suitable for deployment on UAV platforms equipped with edge GPUs, supporting real-time onboard pavement perception in real-world scenarios. The main practical advantage lies in its ability to provide comprehensive pavement condition assessment in one pass, which can support maintenance prioritization and automated reporting systems. Moreover, the proposed framework can be extended to vehicle-based pavement inspection and other multi-task perception applications.