Research On Front Vehicle Detection Method Based On Fusion Of Millimeter Wave Radar And Machine Vision

The front vehicle is an important part of the automatic driving perception system to recognize the traffic flow scene.The research of low-cost and high reliable front vehicle detection method is an important means to improve the safety of intelligent vehicles and promote the commercialization of intelligent driving vehicles.There are many limitations in using a single sensor to realize the front vehicle detection.At present,the multi-sensor fusion method of front vehicle detection is the focus of research at home and abroad.By making full use of the sensor characteristics of millimeter wave radar and machine vision,the data fusion is used to realize the intelligent front vehicle detection,so as to ensure the high reliability,high accuracy and high robustness of detection.The existing research mainly realizes the fusion of MMW radar and machine vision by the way of feature level data fusion,and carries on the visual target detection based on the radar detection results,which is greatly affected by the radar detection error and sensor calibration error.Based on the full analysis and summary of the existing methods,this paper proposes a data fusion framework of millimeter wave radar and machine vision object level.Through the design of millimeter wave radar target detection algorithm and machine vision target detection algorithm respectively,the detection process is independent.Finally,the detection results are fused by the target association algorithm.The results show that this method has better performance than the existing methods.First of all,according to the data processing of MMW radar,this paper analyzes the data characteristics of MMW Radar under different working conditions,puts forward a primary target selection algorithm to achieve the radar invalid data and zero data,and puts forward a radar target clustering algorithm based on hierarchical clustering to achieve the separation of dynamic target and static target of radar,effectively eliminating the road guardrail in traffic environment The influence of static obstacles such as traffic signs on the detection system.In addition,this paper proposes a radar target track management method based on finite state machine,which uses Kalman filter to effectively estimate the azimuth information and motion information of radar target,and uses the method of global nearest neighbor association to achieve the matching of up and down frames of radar target,effectively suppresses the interference of radar hop data,and ensures the reliability and robustness of radar detection results Sex and accuracy.Secondly,for the data processing of machine vision,yolov3 deep convolution neural network is used to realize the image target detection.In view of the limitation of mobile vehicle environment on calculation force,this paper proposes the optimization of network model.The back-end of the network is changed from Darknet53 to MobileNetV2 by the method of refined model design.The network is compressed based on sparse training and APoZ model pruning standard,and redundant neurons are eliminated.At the cost of small precision loss,the model volume is effectively reduced and the detection efficiency of the model is greatly improved.The detection speed under CPU environment reaches 70ms FPS.In addition,in order to improve the detection ability of the system for long-distance small target vehicles,this paper proposes a target tracking method based on kernel correlation filtering,which uses multi-scale template and normalized cross-correlation coefficient as similarity measurement index to fuse the network detection results and image tracking results through nearest neighbor association algorithm,so as to solve the scale insensitivity of kernel correlation filtering tracking The shortcomings of the system effectively expand the detection range.Then,aiming at the fusion of millimeter wave radar and machine vision data,this paper realizes the time synchronization of radar data and vision data by using the minimum common sampling frame method,and proposes a method of association between radar data and image data based on direct linear transformation.Using the midpoint of the shadow line of the vehicle bottom as the indirect homonymous point of the radar plane and the image plane,the radar coordinates of the homonymous point are obtained by the initial calibration of the sensor and the projection and association of the radar target to the image based on the coordinate space conversion model.An adaptive thresholding image segmentation method is proposed to obtain the shadow lines of the vehicle bottom based on the analysis of the connectivity of the image spots and the analysis of the minimum circumscribed rectangle.The consistency algorithm of random sampling is used to solve the transformation model,which effectively solves the impact of the same name point mismatch on the accuracy of the model,and ensures the accuracy and reliability of the model.Finally,the real vehicle test platform is built to test the method in the traffic flow scene under different light conditions and different road conditions.According to the verification of MMW radar data processing algorithm,the results show that this method can well screen the dynamic targets in the road environment and effectively estimate the power information and azimuth information,and the tracking process is stable.According to the verification of machine vision data processing algorithm,the results show that under normal light conditions,the method can detect vehicles well.Under extreme light conditions,the method in this paper has a general performance under extreme conditions due to the lack of extreme samples in the training process.In addition,the image tracking method ensures the detection of long-distance targets Quality measurement.According to the verification of millimeter wave radar and visual data fusion method,the results show that the method in this paper has good performance in actual measurement,and the detection range is more than 150 meters.By comparing this method with the existing method in the same experimental environment.The results show that the object level target fusion method in this paper has higher stability,accuracy and robustness than the existing feature level target fusion method in the actual traffic scene detection.