Research On Vehicle Target Detection Based On Information Fusion Of Roadside Millimeter Wave Radar And Camera

With the development of V2 X vehicle network technology,vehicle-road collaboration technology has become a research hotspot.Target detection sensors installed on the side of the road can provide a more low-cost,high-real-time and high-accuracy data source for vehicle statistics,collision prevention warning,real-time scheduling,and reduce the dependence on single vehicle intelligence.At present,the mainstream target detection sensors are millimeter-wave radar and camera.The millimeter-wave radar can detect the distance and speed of the target,but it can not accurately identify the target type,and it is easy to generate false targets by noise interference.Cameras can detect a wide range and obtain rich image information,but it requires high environmental factors.Therefore,aiming at the shortcomings of millimeter wave radar and camera,this thesis designs a road-side vehicle target detection model based on the information fusion of millimeter wave radar and camera,so as to realize information complementarity and improve detection accuracy and real-time performance.The main work of this thesis includes the following aspects:1.Firstly,this thesis analyzes the necessity of vehicle target detection and the performance of different sensors in target detection,summarizes the research status and main problems of vehicle target detection algorithm and multi-sensor information fusion,and designs the architecture of vehicle-road cooperation road-side target detection system.2.In view of the situation that the millimeter-wave radar is susceptible to the influence of the environment to produce false targets or even misdetect,the target detection range threshold is set by combining C-V2 X high-precision positioning information to select the target within the effective range in this thesis,and the filter algorithm is designed to preprocess the millimeter-wave radar data to improve the reliability and validity of the data.3.In view of the lack of real-time performance of camera detection,the thesis designed a detection algorithm based on YOLO deep learning in the Darknet framework,optimized the structure of multi-scale feature fusion network,trained the network with object labels to obtain weight files,and improved the performance of small target detection.4.This thesis designs the fusion framework of millimeter wave radar and camera,establishes the time and space joint calibration model of millimeter-wave radar and camera,and realizes the data association algorithm based on target ROI region.At the same time,aiming at the large detection error caused by the change of sensor detection accuracy in different scenes,this paper designs an adaptive target detection algorithm based on interactive multi-model,which keeps the system with high detection accuracy by changing the weight of model.5.Finally,through the test,the fusion method designed in this thesis can accurately detect the target vehicle,the identification accuracy can reach 92%,and has good real-time performance,the Jetson AGX Xavier server detection rate can reach 30 FPS,and this method is able to adapt to different scenes,the system stability is good,which can provide more reliable traffic,road,vehicle information for Vehicle-road coordination platform.