With the increasing demand for perception in complex road environments in intelligent driving, rapid and accurate identification of paved-road obstacles has become a critical prerequisite for driving safety and comfort. Various types of road obstacles can significantly affect vehicle stability and ride quality. To address this challenge, a lightweight and efficient vision-based obstacle detection framework, termed GroupScale-YOLO, is proposed, in which detection accuracy and computational efficiency are jointly enhanced through the collaborative design of multiple novel modules. First, a dedicated dataset targeting common paved-road obstacles is constructed, and six data augmentation strategies are employed to mitigate the adverse effects of road surface undulations and illumination variations on visual perception. Second, to overcome the limitations of YOLOv11n in paved-road obstacle detection tasks, targeted optimizations are introduced to the backbone network, convolutional blocks, and detection head. Experimental results indicate that GroupScale-YOLO achieves a 29.95% reduction in model parameters while simultaneously increasing mAP@0.5 by 0.6% on the self-built dataset, demonstrating its suitability for deployment in resource-constrained scenarios. Furthermore, real-vehicle road tests confirm that the proposed method maintains stable and accurate obstacle detection performance under practical driving conditions, offering a reliable solution for intelligent vehicle environmental perception.

Road sign detection is crucial for highway maintenance but often suffers from sign loss, occlusion, and spatial misjudgments such as repeated local detections or mapping errors. To address these issues, this study proposes YOLO-DeepSort, a tracking and detection framework integrating a novel Spatial Multivariate Clustering Algorithm with GPS information (SCMCg). The YOLOv9 detector is augmented using Mixed Local Channel Attention (MLCA) and DualConv modules to enhance image feature extraction and contextual awareness while compressing the theoretical model volume. In the tracking phase, DeepSort combined with SCMCg employs Delaunay triangulation and hierarchical GPS constraints to refine spatial clustering and data association. Validation was conducted on a mixed dataset comprising the CCTSDB and self-collected images from a Ningxia national road. Experimental results indicate that the proposed model operates efficiently at 21.4 M parameters, achieving a precision of 97.8%, a mean Average Precision (mAP) of 91.2%, and a tracking Success Rate of 97.3%. Compared to the baseline YOLOv8-DeepSort, absolute improvements of 4.60% in precision and 4.20% in mAP were observed. The integrated framework effectively mitigates occlusion and tracking spatial errors, providing a robust and lightweight methodology for the automated condition assessment of intelligent transportation infrastructure.