| The environment sensing system is one of the most essential systems for intelligent driving vehicles.The road environment where vehicles drive is complex and variable,and also the perception system is easily affected by interference factors such as bad weather and complex lighting.In order to provide a high-resolution,low-cost and highly reliable perception system for autonomous vehicles,this paper takes automotive millimeter radar perception technology as the main research line,aims to improve the performance of intelligent perception system,deeply studies the multi-target detection and tracking method of traditional millimeter wave radar and camera fusion perception,focuses on exploring the application prospects of new generation 4D imaging millimeter wave radar in traffic participant classification and3 D detection.The core research of this paper mainly includes:(1)A target detection algorithm with traditional millimeter-wave radar and camera feature-level fusion is proposed.Firstly,mapping the millimeter-wave radar point cloud into the image coordinate system through the transformation matrix constructs the radar feature map.Then,introducing convolutional attention to construct a radar attention module.Next,filtering radar attention module into detection backbone network constructs a feature-level fusion network.This network integrates radar point cloud and image feature information,which can improve the probability of target detection under complex background interference.The validation results on the nu Scenes dataset show that the average detection accuracy of the feature-level fusion network is 5.57% better than that of the single-modality image detection network.The validation results of the scene data collected in a real vehicle at night,rain,fog and hazy weather also show that the fusion network can effectively improve the target detection accuracy in complex scenes such as low light and bad weather.(2)A target tracking algorithm for traditional millimeter-wave radar and camera fusion is proposed.Firstly,the detection results of millimeter wave radar and camera are correlated in the image plane to generate a random finite set with target types.Then,based on the Gaussian mixed probability hypothesis density filter framework,the measurement loss labels,elliptic discriminant threshold and decay function are introduced to maintain targets’ life cycle,and the pruning and merging processes are improved to reduce the complexity of the algorithm.The experimental results show that the tracking algorithm proposed in this paper can accurately estimate the number and state of targets in the cases of target occlusion and measurement loss,which improves the target tracking stability and trajectory continuity(3)A traffic participant classification algorithm based on 4D imaging millimeter wave radar is proposed.Firstly,the 4D imaging millimeter wave radar development board was used to collect point cloud data of five types of typical traffic participants:pedestrians,bicyclists,motorcyclists,cars and large buses in dynamic and static scenes,and 10,000 data samples were produced.Then,a machine learning-based classifier was proposed based on the manual feature extraction method.A point cloud feature extraction model was designed to extract a total of 45-dimensional feature vectors;the optimal 20-dimensional feature vectors were selected by the feature filtering method;the experimental results showed that the average classification accuracy of the machine learning classifier was 92% when the targets were classified into five categories,and the average classification accuracy could reach more than95% when they were classified into four categories.Then,the point cloud classification network Radar Transformer is proposed based on the neural network feature extraction method,which takes the attention mechanism as the core and adopts a combination of scalar attention and vector attention to make full use of the spatial,Doppler velocity and reflection intensity information of radar point clouds to achieve the deep fusion of local features and global features.With the same test set,Radar Transformer achieves 94.9% accuracy in five classifications,which is better than traditional machine learning methods.This study experimentally demonstrates the effectiveness of next-generation 4D imaging millimeter wave radar in traffic participant classification,solving the problem of poor classification performance of traditional millimeter wave radar.(4)A 3D target detection algorithm based on 4D imaging millimeter wave radar is proposed.Firstly,a data acquisition platform was built using 4D imaging millimeter wave radar,32-line Li DAR,monocular camera and high-precision positioning navigation system.Then,the synchronization data of various road scenes at different time periods were collected and the dataset TJ4 DRad Set was produced with the LIDAR results as the true value,containing a total of 7757 synchronization frames for44 scenes.Next,the 4D radar point cloud coding method is reconstructed and coordinate attention is introduced to capture the point cloud features,to build the 3D target detection network PPCANet.in the test set,the single-frame detection result is35.86 for 3DmAP and 40.79 for BEVmAP,and the detection result is 42.61 for3 DmAP and 49.07 for BEVmAP after 4 frames of point cloud superposition.The average detection accuracy of PPCANet is better than other comparison algorithms,while achieving the balance of detection accuracy and processing speed.The results show that the proposed traditional radar and camera fusion algorithm in the paper can improve the reliability and robustness of the perception system.and validates the excellent performance of the next generation 4D imaging millimeter wave in sensing.The research in this paper has important reference significance for the solution design and performance improvement of perception system for autonomous vehicles. |
PDF Full Text Request