| As an important part of intelligent connected vehicles,environmental perception plays an increasingly important role in the safety of autonomous driving.The main purpose of environment perception is to perceive the surrounding complex environment timely and accurately.In order to improve the detection accuracy and detection speed of long-distance and small target vehicles under bad working conditions,an information fusion method based on millimeter wave radar and vision is proposed in this paper.The main research contents are as follows: an important part of intelligent connected vehicles,environmental perception plays an increasingly important role in the safety of autonomous driving.The main purpose of environment perception is to perceive the surrounding complex environment timely and accurately.In order to improve the detection accuracy and detection speed of long-distance and small target vehicles in the perception system under bad working conditions,this paper proposes an information fusion method based on millimeter wave radar and vision for vehicle detection.The main research areas are as follows:（1）After reading a large number of relevant references,the research direction and ideas of this article are clearly based on the current research methods,applicable scenarios and existing shortcomings of vehicle detection.（2）The Continental ARS408-21 millimeter wave radar is used in this article.The information collected by the millimeter wave radar is divided and the reasons for the various signals are summarized.The PSO-SVM algorithm is innovatively proposed to realize the selection of vehicle targets in this paper.The experimental results show that the accuracy of this method to obtain effective targets is as high as 95.3%,compared with the threshold method,the accuracy is increased by 13.45%,and it has a better target selection effect.（3）The visual detection algorithm is improved on the basis of the original YOLOv3-tiny algorithm.In this paper,a deep separable convolution layer is used to replace the traditional convolution layer,which increases the number of feature extraction layers and detection scale.In addition,the SENet attention mechanism is added to the network,which reduces the amount of network calculations and improves the feature extraction ability of key information of the target.Finally,ablation experiments were performed on the three improved measures in the BDD100 K dataset of multi-scene vehicle detection.The results show that the improved methods proposed in this paper have improved the model detection effect.Among them,after changing the traditional convolution to the depth separable convolution and adding the increase of the SENet module,the m AP increased by 8.36% and 11.94%,respectively.（4）In order to improve the detection accuracy and speed of small target vehicles under bad working conditions,the vehicles are detected in this paper based on the fusion of millimeter wave radar and visual information.Firstly,the millimeter-wave radar is installed and calibrated;secondly,the spatial alignment of the coordinate system is realized by the relationship of rotation and translation,and the parameters of the camera are obtained by the Zhang’s calibration method;then,the time alignment is realized by the thread synchronization method.Finally,according to the fusion rules,the corresponding target category,location and status information are comprehensively output.After verification,the results show that in the daytime scene,the visual detection algorithm detection accuracy rate is 91.5%,and each frame takes 73 ms,and the fusion detection algorithm detection accuracy rate is 93.7%,and each frame takes 70ms;in the night scene,the visual detection algorithm The detection accuracy rate is 88.6%,and each frame takes78 ms.The fusion detection algorithm detection accuracy rate is 92.1%,and each frame takes 73ms;the fusion scheme effectively improves the target detection performance and makes up for the shortcomings of a single camera detection. |
PDF Full Text Request