<p>With the rapid advancement of urbanization and the continuous increase in vehicle ownership, traffic congestion and frequent accidents have become pressing challenges that demand urgent solutions. In this context, breakthroughs in intelligent vehicle technology, particularly in the field of forward visual perception, are especially critical. Traffic scene instance segmentation requires a balance between high precision and real-time performance in complex and dynamic environments, yet existing technologies exhibit significant limitations. This study aims to address the imbalance among accuracy, efficiency, and adaptability in traffic scene instance segmentation. To this end, we propose UDS-YOLO: an improved network based on YOLOv8-seg, which integrates three key components—the UniRepLknet feature extraction module for enhanced spatial feature capture, the Dysample dynamic upsampling technique for adaptive detail recovery, and the parameter-free SimAM attention mechanism for focused feature weighting. Experimental results on the Cityscapes dataset demonstrate that UDS-YOLO achieves a mAP@0.5 of 39.5% and a mAP@0.5:0.95 of 24%, outperforming mainstream models such as YOLOv5-seg and YOLOv8-seg, while maintaining a frame rate (FPS) of 106.6. The network requires only 11.4 GFLOPs, ensuring high efficiency and suitability for resource-constrained environments. Through the synergistic optimization of ‘feature extraction—upsampling—attention mechanism,’ it bridges the gap in ‘accuracy—efficiency—complex scene adaptability’ present in existing methods. This study provides a robust solution for real-time instance segmentation in autonomous driving, balancing lightweight design with high-precision requirements to meet the demands of complex traffic scenarios. The research offers a practical and feasible technical solution for instance segmentation tasks in autonomous driving systems, with broad application prospects.</p>