摘要
为尝试利用仿真平台生成的数据集来构建自动驾驶场景下的目标检测模型,探索仿真技术与实际应用的融合路径,基于Carla模拟环境平台,通过配置多种交通工具、行人、动物等元素,生成包含多类目标的图像数据集并进行详细的标注处理,利用改进的YOLOv5目标检测算法来训练模型,比较不同超参数设置下模型的检测性能。结果表明,仅依靠仿真平台生成的数据进行训练的YOLOv5模型在验证集上的平均精度(mAP)达到88%,该结果验证了构建的基于仿真技术的目标检测框架是有效的。进一步分析认为,仿真技术在目标检测领域具有重要优势,可生成大量包含各类目标的图像,有助于目标检测模型的训练,为后续利用迁移学习将仿真环境中的知识迁移到实际应用中及探索仿真与实际应用融合的路径奠定了基础。
In order to try to use the data set generated by the simulation platform to build the target detection model in the automatic driving scene,the study explores the fusion path of simulation technology and practical application,generates the image data set containing multiple types of targets and marks them in detail based on the Carla simulation environment platform by deploying various traffic vehicles,pedestrians,animals,etc.The improved YOLOv5 target detection algorithm is used to train the model,and the detection performance of the model under different hyperparameter settings is compared.The results show that the YOLOv5 model trained only on the data generated by the simulation platform has an average accuracy(mAP)of 88% on the verification set.The results verify that the object detection framework based on simulation technology is effective.Further analysis shows that simulation technology has important advantages in the field of target detection,and can generate a large number of images containing various targets,which is conducive to the training of target detection model,and lays a foundation for the subsequent use of transfer learning to transfer the knowledge in the simulation environment to the practical application,and explore the path of the integration of simulation and practical application.
作者
尹誉翔
Yin Yuxiang(School of Traffic and Transportation,Lanzhou Jiaotong University,Lanzhou 730070,China)
出处
《黑龙江科学》
2024年第6期12-15,共4页
Heilongjiang Science
