摘要
针对自动驾驶领域远距离目标的漏检问题,提出一种融合相机与毫米波雷达数据的改进CenterFusion的3D目标检测模型。首先,引入早期融合策略将雷达数据映射到图像平面上,并将其与图像数据结合形成多通道输入,以增强网络模型的抗干扰能力。其次,在特征融合网络后引入注意力机制,使模型聚焦于融合特征图关键信息提取,有效提高了3D目标检测的准确度。然后,进一步改进损失函数解决正负样本不均衡问题。最终,模型在NuScenes数据集上进行对比实验和消融实验。结果表明,改进模型相较于传统的CenterFusion模型平均检测精度提高了1.5%,NuScenes检测分数提高了2.1%,有效提高了远距离目标的检测能力。
Aiming at the problem of missed detection of long-distance targets in the field of autonomous driving,an improved 3D target detection model based on CenterFusion is proposed,which combines camera and millimeter wave radar data.First of all,the early fusion strategy is introduced to map the radar data to the image plane and combine it with the image data to form multi-channel input to enhance the anti-jamming ability of the network model.Secondly,after the feature fusion network,the attention mechanism is introduced to make the model focus on the key information extraction of the fusion feature map,which effectively improves the accuracy of 3D target detection.Then,the loss function is further improved to solve the problem of imbalance between positive and negative samples.Finally,the proposed model is used to carry out comparative experiments and ablation experiments on nuScenes data sets,and the results show that the average detection accuracy of the improved model is 1.5%higher than that of the traditional CenterFusion model,and the NuScenes detection score of the improved model is 2.1%higher,effectively improving the detection ability of long-distance targets.
作者
曾恒
姚娅川
ZENG Heng;YAO Yachuan(School of Automation and Information Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;School of Physics and Electronic Engineering,Sichuan University of Science&Engineering,Yibin 644000,China;Intelligent Perception and Control Key Laboratory of Sichuan Province,Yibin 644000,China)
出处
《四川轻化工大学学报(自然科学版)》
2025年第4期48-57,共10页
Journal of Sichuan University of Science & Engineering(Natural Science Edition)
基金
四川省科技厅重大专题项目(2018GZDZX0045)。
