基于深度学习与数据回归的智能车单目测距方法  

Intelligent vehicle monocular ranging method based on deep learning and data regression

作　　者：郭贺 张蕊[1,2] 成英 彭涛[1,2] 夏鹏 张扬 GUO He;ZHANG Rui;CHENG Ying;PENG Tao;XIA Peng;ZHANG Yang(School of Automobile and Transportation,Tianjin University of Technology and Education,Tianjin 300222,China;National and Local Joint Engineering Research Center for Intelligent Vehicle Infrastructure Cooperation and Safety Technology,Tianjin 300222,China;Tianjin Dolphin Zhixing Technology Co.,Ltd.,Tianjin 300380,China)

机构地区：[1]天津职业技术师范大学汽车与交通学院,天津300222 [2]智能车路协同与安全技术国家地方联合工程研究中心,天津300222 [3]天津海豚智行科技有限公司,天津300380

出　　处：《公路交通科技》2025年第3期11-20,共10页Journal of Highway and Transportation Research and Development

基　　金：国家自然科学基金项目(52172350);天津市教委科研计划项目(2020KJ120,2020KJ122);天津市应用基础研究多元投入基金项目(21JCZDJC00700)。

摘　　要：【目标】针对自动驾驶中单目测距精度不高与泛用性不强的问题,提出了一种基于深度学习与数据回归的单目测距方法。【方法】提出了一种改进的YOLOv5目标检测算法。首先,采用K-means++算法对先验框进行重新聚类,优化初始聚类中心点。其次,在主干网络中引入注意力机制,有效提升特征提取能力。同时,将原损失函数替换为E-IOU损失函数,显著降低了目标定位误差。然后,基于改良的目标检测算法,构建了单目测距模型。该模型依托于不同检测对象的斜角长度与摄像头的真实距离建立了若干回归方程,扩充了单位图像内的像素点数目,提高了检测精度。最后,该模型利用卡尔曼滤波融合了汽车的运动数据,实现了对检测距离的双重数据融合校正。【结果】改进后的算法检测精度较原始YOLOv5提升0.68%,误报率得到明显改善。本研究算法在10~100 m内的测量平均相对误差为3.67%,在30~100 m内的测量平均相对误差为3.24%。通过英伟达加速推理器TensorRT对整套算法进行加速,单张图片推理速度达到12 ms,比未加速的算法处理速度提升了5倍,比双目算法处理速度提升了4倍,且在大于30 m的测距情景中可以取得最优效果。【结论】[HTK]本研究提出算法可以满足一般精度要求的智能车辆与非智能车辆的测距任务。[Objective]In view of the problems of low accuracy and poor generality of monocular ranging in autonomous driving,a monocular ranging method based on deep learning and regression was proposed.[Method]An improved YOLOv5 object detection algorithm was proposed.First,the K-means++algorithm was used to re-cluster the prior boxes,optimizing the initial cluster centers.Second,an attention mechanism was introduced into the backbone network,effectively enhancing the feature extraction capabilities.Meanwhile,the original loss function was replaced with the E-IOU loss function,significantly reducing the target localization errors.Then,the monocular ranging model was constructed based on the improved object detection algorithm.The model established several regression equations based on the diagonal lengths of different detected objects and their real distances from camera,increasing the number of pixels per unit image,and enhancing the detection precision.Finally,the model used Kalman filtering to integrate vehicle motion data,achieving the dual-data fusion correction for ranging.[Result]The detection accuracy of improved algorithm increases by 0.68%compared with the original YOLOv5 with noticeable reduction in false positives.The average relative error of proposed algorithm is 3.67%within the range of 10-100 m,and 3.24%within the range of 30-100 m.The entire algorithm is accelerated by using NVIDIA[DK]’s TensorRT.The inference speed for a single image reaches 12 ms.That is 5 times faster than that with the unaccelerated version,and 4 times faster than that with the binocular algorithm.The proposed method achieves optimal performance in ranging scenarios beyond 30 m.[Conclusion]The proposed algorithm can meet the ranging requirements for both intelligent and non-intelligent vehicles with general accuracy needs.

关 键 词：智能交通 单目测距 深度学习 注意力机制 数据回归 

分 类 号：U463.6[机械工程—车辆工程]