光斑密度峰值的毫米波雷达目标检测

Millimeter-wave Radar Object Detection Based on Spot Density Peak

毫米波雷达目标检测任务是车辆环境感知的重要组成部分,对恶劣天气条件下的智能驾驶具有重要意义.虽然现有雷达目标检测方法已取得不错的研究成果,但雷达数据仍存在手工设计的特征信息量不足、特征提取不充分以及时序特征未充分利用的问题.为解决这些问题,本文提出光斑密度峰值与神经网络相融合的两阶段目标检测方法.第一个阶段光斑密度峰值聚类算法,对射频图像中的目标进行粗略估计,并生成聚类簇对应的目标候选.将聚类生成的候选目标特征融合到原始射频图像.第二阶段基于通道融合的3D自编码器目标检测网络进一步提取目标多普勒速度和时序特征并分类.实验表明,所提出的两阶段方法与基准实验RODNet(CDC)相比,平均精度指标提升4.3%,平均召回率提升2.3%.

The task of object detection using millimeter-wave radar is an important part of vehicle environment perception and has significant implications for intelligent driving under adverse weather conditions.Although existing radar-based object detection methods have achieved promising results,there are still issues related to insufficient manually designed feature information,inadequate feature extraction,and untapped temporal features.To address these issues,this paper proposes a two-stage object detection method based on density peak of the spot and neural network.In the first stage,a spot density peak clustering algorithm is used to provide a rough estimate of the target in the RF image and generate target candidates corresponding to the clustering clusters.The features of these clustering-generated target candidates are then fused into the original RF image.In the second stage,a 3D autoencoder-based object detection network using channel fusion is used to further extract target Doppler velocity and temporal features and classify them.Experimental results demonstrate that the proposed two-stage method outperforms the benchmark experiment RODNet(CDC)in terms of average precision and average recall rate,with an improvement of 4.3%and 2.3%,respectively.