| Self-driving cars can improve the active safety performance and ease traffic congestion,which is the future development direction of traffic.Perception system is the basis for selfdriving.A variety of sensors are used to obtain information about the surrounding environment.Only when accurate and reliable perception information is input to the decision-making system can get appropriate decision information.Vehicles are the most common and most prone to collision risk.In this paper,the vehicle detection and tracking method of millimeter-wave radar camera fusion is studied.The main work is as follows:(1)Through a review of literature,the paper provides an analysis of the existing vehicle detec-tion methods,a comparison of the advantages and disadvantages of common sensors and a summary of the shortcomings of a single sensor and existing fusion methods to detect vehicles.Based on those,the sensor selection and fusion scheme are proposed.(2)The vehicle detection method of millimeter wave radar is explained.The principle and performance of millimeter wave radar are introduced,choose the radar suitable for this work.The radar data and the cause for the error detection are analyzed.A target detection network R-Dense Net based on multi-dimensional data of the millimeter wave radar is proposed,which leads to the detection result of millimeter wave radar.After testing,the detection accuracy and speed of the network meet the subsequent fusion requirements.(3)The vehicle detection method of camera is explained.The common visual recognition networks are introduced,and YOLOv4-tiny network is selected in this work due to its realtime performance,whose principle and architecture are introduced.Vehicle images are collected as self-built data set which is used to train and test the network where the transfer learning is used in training.The experiment was carried out,and the results show that the accuracy and speed of YOLOv4-tiny meet the fusion requirements.(4)A target detection and tracking method based on parallel fusion of millimeter-wave radar and camera is proposed.Firstly,the realization of spatial fusion and time fusion is elaborated.After data level fusion is realized,the decision level fusion method is introduced.The original target points are fused by the fusion rule,this step is to filter out the radar error detection points and noise points.Then the successful fusion points are continu-ously tracked by using the real target survival judgment method.During the tracking process,false detection and missed detection of millimeter wave radar of camera are captured and the lost data are supplemented by Kalman filtering,YOLO re-identification and other methods.(5)The experiments are carried out to verify the effectiveness of the proposed algorithm.Build hard-ware and software experimental platforms to carry out experiments and analyze the results of the experiments.Comparing the proposesd algorithm with the existing algorithms to evaluate the overall accuracy rate,the results show that the accuracy rate of the proposed algorithm increases 2.17% under the premise of real-time performance requirement.The single-sensor detection method and the fusion detection method proposed are compared in cases where the camera missing detection,millimeter wave radar error detection,millimeter wave radar missing detection,and two sensors missing.The analysis results prove that the proposed algorithm based on the parallel fusion is better in terms of robustness and stable tracking. |
