Research On Vehicle Target Lightweight Detection Network Based On MMW Radar Image

The research on vehicle target lightweight detection network based on Two Dimensional Millimeter Wave Radar Occupancy Grid Image(2DMMWROGI)can improve the perception ability of disaster emergency rescue personnel in enclosed spaces such as long tunnels when disasters occur,accelerate rescue speed,and ensure the safety of the affected population.Millimeter wave radar based on motion carriers can achieve self-localization in enclosed spaces through simultaneous localization and mapping(SLAM)technology,and can penetrate smoke barriers,fire barriers,and hightemperature fields to construct radar occupied raster images.This image has more significant target contour semantic features compared to a single frame radar point cloud,which can assist in solving target detection problems in enclosed space fire scenes.In this paper,a lightweight vehicle target detection network based on millimeter wave radar images is studied,which can be deployed on a low-cost edge computing platform.The vehicle target feature in the 2D millimeter wave radar occupied grid image is mainly the dihedral angle scattering formed by the target side and the ground,which is quite different from the optical image of the vehicle.The lightweight vehicle target detection algorithm represented by YOLOv5 s mainly processes optical images,and there is no baseline model for target detection in millimeter wave radar occupied raster images.To solve this problem,this paper uses the transfer learning method,the existing optical image feature extraction ability and a small number of radar image annotation data sets to solve the problem of vehicle target detection in radar image.In addition,this study further optimized the YOLOv5 s model based on the NVIDIA Jetson Xavier NX platform requirements and proposed the YOLOv5s-M(YOLOv5sMobile Net)model.The model uses a MobileNetV3 network to further reduce the number of model parameters,but it leads to a decrease in accuracy.In response to this issue,this paper adopts the K-means clustering algorithm to improve prior box estimation,and utilizes the SENet(Squeeze-and-Excitation Network)attention mechanism to improve model accuracy,completing the construction of the YOLOv5sMKS(YOLOv5s-Mobile Net-K-means-SENet)model.The specific research work is as follows:(1)This paper implements a method for object detection using 2D millimeter wave radar occupying raster maps.First,tag the radar image,and convert it into a VOC format dataset,and complete the data augmentation.Secondly,this paper uses the YOLOv5 s deep learning network as the basic model,which has been pre trained using the VOC dataset of optical images.Finally,through transfer learning on the radar small sample data,the fast convergence and optimization of the model are realized.(2)This paper proposes a YOLOv5s-M network based on MobileNetV3 to address the problem of excessive network parameters in vehicle target detection tasks.Utilize MobileNetV3 to replace some of the backbone networks of YOLOv5 s to reduce computational complexity and parameter count.This paper deploys the network on the low-cost edge computing platform NVIDIA Jetson Xavier NX platform and conducts experiments.The results show that the YOLOv5s-M model in this paper significantly improves the detection speed compared with the original YOLOv5 s,but reduces the accuracy and recall of target detection.(3)This paper proposes an improved YOLOv5s-MKS model based on K-means and SENet to address the issue of reduced detection accuracy and recall caused by simplifying the structure of object detection networks.In this paper,K-means algorithm is used to cluster the tag coordinates of radar image data to obtain a priori frame that is more consistent with the distribution of vehicle targets in 2D radar images.This paper applies the lightweight attention mechanism SENet to object detection networks to enhance the correlation and discrimination between feature channels.This paper conducted experiments on the NVIDIA Jetson Xavier NX platform,and the results showed that compared with YOLOv5s-M,the improved YOLOv5s-MKS significantly improved detection performance and speed while maintaining lower computational costs and parameter count.This method can achieve a test frame rate of 36.29 frames per second,and can recognize stationary vehicles in millimeter wave radar images in real time and accurately.