基于YOLOv7-R的多车辆目标识别  

Multi-vehicle object recognition based on YOLOv7-R

作　　者：李珣[1,2] 伍荣兴 周慧龙 刘欣[3] 高涵 王文杰 Li Xun;Wu Rongxing;Zhou Huilong;Liu Xin;Gao Han;Wang Wenjie(School of Electronics and Information,Xi'an Polytechnic University,Xi'an 710048,China;Xi'an Polytechnic University Branch of Shaanxi Artificial Intelligence Joint Laboratory,Xi'an 710000,China;School of Information Engineering,Xi'an Eurasia University,Xi'an 170068,China)

机构地区：[1]西安工程大学电子信息学院,西安710048 [2]陕西省人工智能联合实验室西安工程大学分部,西安710000 [3]西安欧亚学院信息工程学院,西安170068

出　　处：《东南大学学报（自然科学版）》2024年第5期1260-1270,共11页Journal of Southeast University：Natural Science Edition

基　　金：国家自然科学基金资助项目(51905405);陕西省自然科学基础研究计划资助项目(2022JM407);陕西省教育厅2023年度一般专项科学研究计划资助项目(23JK05777).

摘　　要：针对智能交通控制对于多车辆目标识别高准确率的需求,基于YOLO算法,提出一种改进的YOLOv7-R算法.将全局注意力机制(GAM)引入骨干网络,增强特征提取性能;利用全维动态高效聚合网络(ODEANet)重构主干网络,提高算法的鲁棒性与精度;使用上下文转换器(CoTNet)替换运算量巨大的可扩展高效聚合层网络(E-ELAN),来引导动态注意力矩阵学习,并降低浮点运算量;采用K-means++聚类算法优化先验帧,提高先验帧的匹配度.通过系统性的改进,车辆多目标识别的效率和准确度均得到提升.在自由流、同步流、阻塞流3种交通流下,分别进行了消融实验.结果表明:YOLOv7-R平均识别率分别达到97.13%、94.85%和94.60%,与基线算法相比分别提高了3.65%、3.20%和1.40%;算法的检测帧率分别为74.63、79.37和75.76帧/s.与基线算法相比,YOLOv7-R的浮点运算量降低3.10%,参数量降低13.37%.To meet the requirements of intelligent traffic control for multi-vehicle object recognition with high accuracy,an improved YOLOv7-R(you only look once version 7 for the road)algorithm was proposed based on the YOLO algorithm.The global attention mechanism(GAM)was introduced into the backbone network to enhance feature extraction performance.The omni-dimensional dynamic and efficient aggregation network(ODEANet)was used to reconstruct the backbone network,improving algorithm robustness and accuracy.The computationally intensive extended efficient layer aggregation network(E-ELAN)was replaced with the contextual transformer network(CoTNet),which guides dynamic attention matrix learning and reduces the floatingpoint computation.Additionally,the K-means++clustering algorithm was also used to optimize the prior frame and improve the matching degree.Through systematic improvement,the efficiency and accuracy of multiobject recognition for vehicles were improved.Experiments were conducted based on three traffic flows,namely,free flow,synchronized flow,and blocked flow.The results show that the YOLOv7-R achieves an average recognition accuracy of 97.13%,94.85%,and 94.60%,respectively,which are 3.65%,3.20%,and 1.40%higher than those of the baseline algorithm.Additionally,the detection frames of the algorithm are 74.63,79.37 and 75.76 frames/s,respectively.Compared with the baseline algorithm,the giga floating-point operations per second(GFLOPS)of YOLOv7-R is reduced by 3.10%,and the number of parameters is reduced by 13.37%.

关 键 词：多目标识别 YOLOv7 车辆检测 注意力机制 特征提取 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]