| 摘 要: 随着城市化加速与车辆数量激增,室内停车场车位检测的准确性与实时性成为智能管理的关键挑战。传统视觉方法在光照不均、遮挡及远距离目标易失真等复杂场景下表现不佳。提出改进型目标检测模型 ADM-YOLOv8,在 YOLOv8的基础上引入 ADown模块以增强小目标特征保持,采用 C2f_Dual结构提升特征多样性,并融合 MLLAttention机制强化空间感知与上下文理解。在自建室内车位数据集上的实验结果表明,ADM-YOLOv8在mAP@0.5上较YOLOv8提升2.9个百分点,在mAP@0.5∶0.95上提升8.7%,同时 GFLOPs降低11.0%。结果验证了该模型在提升精度的同时能兼顾效率,为智能交通系统中的车位检测提供了高效可行的解决方案。 |
| 关键词: 室内停车位检测 YOLOv8 轻量化网络 多尺度特征融合 注意力机制 |
|
中图分类号: TP391
文献标识码: A
|
|
| ADM-YOLOv8:Indoor Parking Slot Detection in Complex Scenes |
|
GUO Jun12, ZHANG Shilong1, GUO Jinting1, SHAO Mengzhen1
|
(1.School of Automotive Engineering, Yancheng Institute of Technology, Yancheng 224051, China;
2.School of Automotive Engineering, Nantong Institute of Technology, Nantong 226002, China)
gj_njau@163.com; 2843720058@qq.com; GJTyq965@163.com; shaomzz@163.com
|
| Abstract: With accelerated urbanization and the surge in vehicle numbers, the accuracy and rea-l time performance of parking slot detection in indoor parking slot have become critical challenges for intelligent management. Traditional vision-based methods perform inadequately in complex scenarios such as uneven illumination,occlusion, and distant targets. This paper proposes an improved object detection model, ADM-YOLOv8, which introduces the ADown module based on YOLOv8 to enhance feature preservation for small targets, adopts the C2f_Dual structure to improve feature diversity, and integrates the MLLAttention mechanism to strengthen spatial awareness and contextual understanding. Experiments results on a sel-f built indoor parking lot dataset demonstrate that ADM-YOLOv8 achieves a 2.9 percentage points improvement in mAP@0.5 and an 8.7% improvement in mAP@ 0.5∶0.95 compared to YOLOv8, while reducing GFLOPs by 11.0% . The results validate that the proposed method enhances accuracy while maintaining efficiency, providing an effective and feasible solution for parking slot detection in intelligent transportation systems. |
| Keywords: indoor parking slot detection YOLOv8 lightweight network mult-i scale feature fusion attention mechanism |
