%0 Journal Article %T 基于OSTD-YOLO复杂路况下的障碍物检测研究
Research on Obstacle Detection in Complex Road Conditions Based on OSTD-YOLO %A 赵晓卓 %A 张成涛 %A 许季何 %A 骆远鹏 %A 黎俊宏 %J Journal of Sensor Technology and Application %P 229-241 %@ 2331-0243 %D 2025 %I Hans Publishing %R 10.12677/jsta.2025.133023 %X 随着自动驾驶技术的不断推进，在面对含有小目标、遮挡目标以及目标尺寸变化较大的复杂路况时，障碍物检测中的漏检与误检问题成为亟待攻克的关键难题。为此，设计了名为OSTD–YOLO (Occlusion and Small Obstacle Detection)的算法。该算法以YOLO11n作为基础模型展开构建，主要有两方面的改进。一是构建了增强小目标特征金字塔模块EMTFP (Enhanced Micro-Target Feature Pyramid)，此模块依托PAFPN架构，引入SPDConv对P2特征层进行精细化处理，有效提升了小目标特征提取能力。同时，受CSP思想和OmniKernel启发，研发出CSP-OmniKernel模块，用于整合不同层级提取的各类特征，让特征信息更加精炼。二是在模型输出端引入DyHead注意力机制检测头，实现了对空间、尺度和任务维度的自适应特征增强，全面提升了模型在复杂路况下对障碍物的感知精度。在公开数据集BDD 100k上进行测试后发现，相较于原始的YOLO11n模型，OSTD–YOLO算法的mAP50提升了近5个百分点，参数量为3.58 M，在与其他同类方法的对比试验中展现出参数量与精度的最佳平衡性。一系列试验充分表明，OSTD-YOLO算法能够出色地应对复杂路况下的障碍物检测任务，为自动驾驶技术的落地应用提供了有力支撑。
With the continuous advancement of autonomous driving technology, in the face of complex road conditions with small targets, occluded targets and large changes in target size, the problem of missed detection and false detection in obstacle detection has become a key problem to be solved urgently. To this end, an algorithm named OSTD-YOLO (Occlusion and Small Obstacle Detection) is designed. The algorithm is built with YOLO11n as the basic model, and there are two main improvements. One is to build an Enhanced Micro-Target Feature Pyramid module EMTFP (Enhanced Micro-Target Feature Pyramid). This module relies on the PAFPN architecture and introduces SPDConv to refine the P2 feature layer, which effectively improves the feature extraction ability of small targets. At the same time, inspired by the CSP idea and OmniKernel, the CSP-OmniKernel module is developed to integrate various features extracted at different levels to make the feature information more refined. Second, the DyHead attention mechanism detection head is introduced at the output of the model, which realizes the adaptive feature enhancement of space, scale and task dimensions, and comprehensively improves the model’s perception accuracy of obstacles under complex road conditions. After testing on the public dataset BDD 100 k, it is found that compared with the original YOLO11n model, the mAP50 of the OSTD-YOLO algorithm is improved by nearly 5 percentage points, and the parameter amount is 3.58 M. It shows the best balance of parameter amount and accuracy in the comparison test with other similar methods. A series of experiments fully demonstrate that the OSTD-YOLO algorithm can effectively handle obstacle detection tasks under complex road conditions, providing strong support for the practical application of autonomous driving technology. %K 复杂路况， %K 障碍物检测， %K 深度学习， %K YOLO
Complex Road Conditions %K Obstacle Detection %K Deep Learning %K YOLO %U http://www.hanspub.org/journal/PaperInformation.aspx?PaperID=113826