Abstract: An improved SAF-FCOS radar-vision fusion target detection network is proposed to address the difficulty in effectively utilizing radar point cloud information and image features, as well as the issue of false or missed detections in harsh weather environments. The backbone network structure of SAF-FCOS is improved, and multi-scale fusion of radar feature information is carried out at C3 and C4 feature layers, so that the network model can make full use of radar information. Using improved LNblock module—LNblcok_GAM before detection layers can extract image features at a lower computational cost while improving the detection performance of the network. In terms of regression loss, the improved CEIOU based on EIOU and GIOU is used to replace the GIOU in the original network, improving the detection accuracy of the network and enhancing the robustness of the model. On the NuScenes dataset, the improved network achieves 70.7% mAP0.5:0.95 and 90.5% AP50, respectively, which are 1.7?percentage points and 0.9?percentage points higher than the original network SAF-FCOS. The missed and false detections are effectively reduced, and the overall detection performance of the improved network is better than other classic visual target detection algorithms.

Key words: target detection, radar-vision fusion, SAF-FCOS network, multi-scale fusion, LNblock_GAM

摘要： 针对雷视特征融合目标检测网络难以有效利用雷达点云信息和图像特征，在恶劣天气环境下，仍然容易出现误检、漏检的问题，提出了一种改进SAF-FCOS雷视融合目标检测网络。对SAF-FCOS的骨干网络结构进行改进，在C3、C4特征层进行雷达特征信息的多尺度融合，使网络模型更充分地利用雷达信息；在检测层前使用改进的LNblock模块——LNblcok_GAM，能够以较低的计算成本提取图像特征的同时提高网络的检测性能；在回归损失方面，使用基于EIOU与GIOU改进的CEIOU替换原网络中的GIOU，提高了网络的检测精度，提升了模型的鲁棒性。在NuScenes数据集上，改进网络的mAP0.5:0.95达到了70.7%，AP50达到了90.5%，比原网络SAF-FCOS分别提高了1.7个百分点和0.9个百分点，漏检、误检的情况得到了有效改善，同时，该改进网络的总体检测效果要优于其他经典的纯视觉目标检测算法。

关键词: 目标检测, 雷视融合, SAF-FCOS网络, 多尺度融合, LNblock_GAM