无线电工程

针对变电站安全监控场景下的行人识别，纯视觉传感器目标检测极易受到光照条件和物体遮挡的影响，存在误检和精度不高等问题，主流雷达视觉融合目标检测网络存在实时性差、精度低的问题，提出了一种以YOLOv5作为主干网络的多尺度雷视融合目标检测算法，实验结果显示该算法在平均精度均值(mean Average Precision, mAP)0.5:0.95和帧率(Frames Per Second, FPS)上均显著优于其他主流目标检测算法。在YOLOv5结构上，额外增添了一条毫米波雷达点云多尺度特征融合分支。雷达点云RGB图像首先通过通道压缩(Channel Block Squeeze, CBS)、跨阶局部网络(Cross Stage Partial Network, CSP)模块提取特征之后，继续通过卷积注意力模块(Convolutional Block Attention Module, CBAM)和最大池化层对雷达点云信息进行不同层次的特征提取，使用空间注意力融合模块的改进对雷达点云特征和视觉特征进行3次多尺度特征融合。实验分析表明，所提算法mAP0.5:0.95比原始YOLOv5网络有显著提升，FPS也远优于目前主流雷视融合算法。

In the context of substation safety monitoring, target detection with pure visual sensor is easily affected by illumination conditions and object occlusion, resulting in problems such as false detection and low accuracy. The mainstream radar vision fusion target detection network has poor real-time performance and low accuracy. A multi-scale radar vision fusion target detection algorithm using YOLOv5 as the backbone network is proposed, the experimental results also show that the algorithm is significantly better than other mainstream target detection algorithms in both mean Average Precision(mAP)0.5:0.95 and Frame Per Second(FPS). On the structure of YOLOv5, an additional branch of millimeter-wave radar point clouds multi-scale feature fusion is added. The radar point clouds RGB image extracts features firstly through the Channel Block Squeeze(CBS) and Cross Stage Partial Network(CSP) modules, and then continues to extract different levels of features from the radar point clouds information through the Convolutional Block Attention Module(CBAM) and maximum pooling layer. The improved spatial attention fusion module is used to perform the three-time multi-scale feature fusion of radar point clouds features and visual features. Experimental analysis shows that the mAP0.5:0.95 of the algorithm proposed is significantly improved compared to the original YOLOv5 network, and the FPS is also far superior to the current mainstream radar vision fusion algorithm.