融合毫米波雷达与机器视觉的雾天车辆检测  

Vehicle detection in foggy weather combining millimeter wave rada and machine vision

作　　者：李颀[1] 叶小敏 冯文斌 LI Qi;YE Xiao-min;FENG Wen-bin(College of Electronic Information and Artificial Intelligence,Shaanxi University of Science and Technology,Xi′an 710021,China)

机构地区：[1]陕西科技大学电子信息与人工智能学院,陕西西安710021

出　　处：《液晶与显示》2023年第10期1445-1454,共10页Chinese Journal of Liquid Crystals and Displays

基　　金：陕西省科技厅项目(No.S2023-YF-YBNY-0232);西安市科技计划(No.22NYYF064);西安市未央区科技计划(No.201305)。

摘　　要：车辆检测对于辅助驾驶系统至关重要,由于雾天道路场景严重退化,图像中的车辆信息不明显,导致车辆检测存在漏检、误检的问题。针对上述问题,本文提出了一种融合毫米波雷达和机器视觉的雾天车辆检测方法。首先,采用暗通道去雾算法对图像进行预处理,提高雾天图像中车辆信息的显著性。然后,采用知识蒸馏改进YOLOv5s算法,在YOLOv5s的特征提取网络中引入知识蒸馏,在目标定位和分类阶段计算蒸馏损失,对损失进行反向传播训练小型网络模型,在保证视觉检测准确度的同时提高检测速度。最后,采用基于潜在目标检测区域搜索的距离匹配算法对视觉检测结果和毫米波雷达检测结果进行决策级融合。以检测目标的类型和距离为匹配依据,滤除干扰信息和错误信息,保留毫米波雷达检测和视觉检测融合后的检测置信度较高的目标,从而提高车辆检测的准确率。实验结果表明,该方法在雾天下最高检测准确率达92.8%,召回率达90.7%,能够实现雾天对车辆的检测。Vehicle detection is very vital to the assisted driving system.Due to the serious degradation of the foggy road scene,the vehicle information in the image is not obvious,resulting in missed detection and false detection problems in vehicle detection.Aiming at the above problems,a vehicle detection method in foggy weather combining millimeter-wave radar and machine vision is proposed.First,the dark channel dehazing algorithm is used to preprocess the image to improve the salience of vehicle information in the image under foggy conditions.Then,the knowledge distillation is used to improve the YOLOv5s algorithm,and the knowledge distillation is introduced into the feature extraction network of YOLOv5s,which is used in the target positioning and classification stages to calculate the distillation loss and backpropagate the loss to train a small network model to improve the detection speed while ensuring the accuracy of visual detection.Finally,the distance matching algorithm based on the search of potential target detection areas is used to compare the visual detection results and the millimeter-wave radar detection results decision-making fusion.Based on the type and distance of the detected target,the interference information and erroneous information is filtered out,and the targets with high confidence after fusion in millimeter-wave radar detection and visual detection is retained.Thereby,the accuracy of vehicle detection is improved.The experimental results show that the method has the highest detection accuracy rate of 92.8%and the recall rate of 90.7%in foggy weather,which can realize the detection of vehicles in foggy weather.

关 键 词：车辆检测 毫米波雷达 去雾 知识蒸馏 距离匹配 决策级融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]