基于轻量化YOLOv5的雷达图像道路目标检测方法

Radar Image Road Target Detection Method Based on Lightweight YOLOv5

针对现有车载毫米波雷达目标检测信号处理流程中未滤除静止杂波干扰和网络参数量及浮点数运算量多的问题,文中提出一种车载毫米波雷达目标检测信号处理链,并基于Ghost卷积和空间深度(Space-to-depth,SPD)卷积,设计了一种车载雷达轻量化YOLOv5目标检测模型。首先,使用平均相消算法对雷达原始AD采样数据进行静止杂波干扰抑制;然后,进行二维快速傅里叶变换(Fast Fourier Transform,FFT)得到目标的距离-多普勒(Range-Doppler,RD)图像;最后,使用轻量化目标检测网络模型对RD图像进行检测。实测数据验证了本文所提方法能够滤除静止杂波,有效减少了目标检测模型的浮点运算量,进而降低了模型的参数量,且具有较高的检测精度。

Addressing the issue of unfiltered stationary clutter interference and the excessive number of network parameters and floating-point operations in the current vehicular millimeter-wave radar target detection signal processing,this paper proposes a signal processing chain for vehicular millimeter-wave radar target detection.This methodology is underpinned by the integration of Ghost convolution and Spatial Depth(SPD)convolution,culminating in the design of a lightweight YOLOv5 target detection model tailored for vehicular radar applications.The process initiates with the suppression of stationary clutter interference through the application of the mean subtraction algorithm on the raw AD sampled data.This is followed by the generation of the target’s Range-Doppler(RD)image through a two-dimensional Fast Fourier Transform(FFT).Subsequently,the lightweight target detection network model is employed for the detection within the RD images.Empirical data corroborates the efficacy of the proposed approach in filtering out stationary clutter,significantly reducing the floating-point computation requirement of the target detection model,thus minimizing the model’s parameter count while simultaneously achieving high detection accuracy.