基于注意力机制的毫米波雷达和视觉融合目标检测算法  

A Millimeter-wave Radar and Vision Fusion Target Detection Algorithm Based on Attention Mechanism

作　　者：陈州全 黄俊[1] 郑元杰 CHEN Zhouquan;HUANG Jun;ZHENG Yuanjie(School of Communication and Information Engineering,Chongqing University of Posts and Telecommunications,Chongqing 400065,China)

机构地区：[1]重庆邮电大学通信与信息工程学院,重庆400065

出　　处：《电讯技术》2023年第10期1574-1581,共8页Telecommunication Engineering

基　　金：国家自然科学基金资助项目(61771085)。

摘　　要：传统目标检测大多基于摄像头采集图像进行,虽然近些年出现了许多优秀的检测网络,但在复杂场景下,仍存在大量漏检、误检等问题。针对这些问题,提出了一种基于注意力机制的毫米波雷达和视觉融合目标检测算法。首先将毫米波雷达数据进行扇形点云柱编码(Fan-shaped Cloud Pillar Code,FCPC),将其转换为前景伪图像;然后,再将其通过坐标关系映射到像素平面,使用卷积注意力模块(Convolutional Block Attention Module,CBAM)对两者特征数据进行融合;采用Yolov4-tiny对融合特征进行检测,并引入Focal Loss对原损失函数进行改进以解决正负样本不均的问题。在Nuscenes数据集上进行模型验证与对比,结果表明,该算法在复杂场景下相比其他单传感器检测算法如Yolov3、Efficientent以及Faster-RCNN等,无论平均检测精度(mean Average Precision,mAP)还是每秒检测帧数(Frames Per Second,FPS)都有明显的提升。Traditional target detection is mostly based on images collected by cameras.Although many excellent detection networks have appeared in recent years,there are still a large number of missed detections and false detections in complex scenarios.For these problems,an attention-based millimeter-wave radar and vision fusion target detection algorithm is proposed.First,the millimeter-wave radar data is converted into a foreground pseudo-image by fan-shaped cloud pillar code(FCPC),and then mapped to the pixel plane through the coordinate relationship,and the Convolutional Block Attention Module(CBAM)is used to fuse the two feature data.Yolov4-tiny is used to detect the fusion features,and Focal Loss is introduced to improve the original loss function to solve the problem of uneven positive and negative samples.Model verification and comparison on the Nuscenes dataset show that the algorithm is better than other single-sensor detection algorithms such as Yolov3,Efficientent,and Faster-RCNN in complex scenes,both the mean Average Precision(mAP)and the number of Frames Per Second(FPS)are significantly improved.

关 键 词：目标检测 视觉融合 数据融合 注意力机制 毫米波雷达 

分 类 号：TN951[电子电信—信号与信息处理] TP391[电子电信—信息与通信工程]