GEIT-YOLO模型在智能驾驶中的研究与应用
Research and Application of GEIT-YOLO Model in Intelligent Driving
为解决传统YOLO模型在复杂道路行驶环境中性能差等问题,提出一种基于改进YOLO11n的智能驾驶目标检测模型GEIT-YOLO。为了让边缘信息融合到各个尺度所提取的特征中,提出GEIT模块,其能够把浅层特征中提取到的边缘信息传递到整个backbone上,并与不同尺度的特征进行融合;同时引入门控注意力机制CGLU,通过门控特征调制联合建模通道依赖与空间上下文关系;最后引入LQE检测头通过显式地评估每个边界框的定位质量,优化NMS阶段并且优先选择更高质量的边界框。实验结果表明:与YOLO11n相比,本文所提出的模型GEIT-YOLO在Udacity数据集上Precision、Recall、mAP50、mAP50:95分别提升了5.6%、2.7%、4.2%、2.3%,改进模型在复杂驾驶环境中的表现更加优异,能显著提升模型检测能力。
To address issues such as the poor performance of traditional YOLO models in complex road environments, this paper proposes GEIT-YOLO, an object detection model for intelligent driving based on an improved YOLO11n. To enable the integration of edge information into features extracted at various scales, the GEIT module is proposed. It propagates the edge information extracted from shallow features throughout the entire backbone network and fuses it with features at different scales. Concurrently, the CGLU gated attention mechanism is introduced to jointly model channel dependencies and spatial contextual relationships through gated feature modulation. Finally, the LQE detection head is incorporated to explicitly evaluate the localization quality of each bounding box, thereby optimizing the Non-Maximum Suppression (NMS) stage and prioritizing bounding boxes of higher quality. Experimental results show that, compared to YOLO11n, the proposed GEIT-YOLO model achieves improvements of 5.6% in Precision, 2.7% in Recall, 4.2% in mAP50, and 2.3% in mAP50:95 on the Udacity dataset. The improved model demonstrates superior performance in complex driving environments and can significantly enhance detection capability.
| [1] |
武彪, 任洪泽, 郑联庆, |
| [2] |
严林志, 周瑛. 从社会诉求到政策因应: 基于文本挖掘的自动驾驶技术发展策略研究[J]. 情报理论与实践, 2025, 48(8): 73-82, 92. |
| [3] |
董甲东, 桑飞虎, 郭庆虎, |
| [4] |
王宁, 智敏. 深度学习下的单阶段通用目标检测算法研究综述[J]. 计算机科学与探索, 2025, 19(5): 1115-1140. |
| [5] |
段立娟, 张子晨, 张广勇. 自监督的两阶段广义小样本目标检测算法[J]. 信号处理, 2025, 41(2): 370-381. |
| [6] |
李威. 面向智能驾驶的交通场景理解方法研究[D]. 长春: 吉林大学, 2021. |
| [7] |
陈钧, 周井泉, 张志鹏. 面向小目标的自校正YOLOv5检测增强算法[J]. 计算机技术与发展, 2024, 34(10): 140-147. |
| [8] |
龙小羽, 南新元. 智能驾驶场景下的中小型障碍物检测方法[J]. 科学技术与工程, 2025, 25(9): 3778-3787. |
| [9] |
董一兵, 曾辉, 李建科, |
| [10] |
徐彦威, 李军, 董元方, |
| [11] |
|
| [12] |
董华军,陈强,李籽骁, |
| [13] |
傅鹏飞, 徐巍, 刘怀广, |
| [14] |
|
| [15] |
|
| [16] |
王坤, 冯康威. FFE-YOLOv8: 交通场景遮挡目标检测算法[J]. 计算机工程与应用, 2025, 61(15): 156-166. |
安徽省高校自然科学研究重点项目(2022AH052821)
/
| 〈 |
|
〉 |