无人驾驶方程式赛车轻量化目标检测算法  

Lightweight Target Detection Algorithm for Unmanned Formula Racing

作　　者：汤文靖 兰建平 周海鹰 Tang Wenjing;Lan Jianping;Zhou Haiying(Institute of Automotive Engineers,Hubei University of Automotive Technology,Shiyan 442002,China;Dongfeng Electronic Technology Co.Ltd,Shanghai 200063,China)

机构地区：[1]湖北汽车工业学院汽车工程师学院,湖北十堰442002 [2]东风电子科技股份有限公司,上海200063

出　　处：《湖北汽车工业学院学报》2023年第2期17-21,27,共6页Journal of Hubei University Of Automotive Technology

基　　金：湖北省教育厅青年人才基金(Q20151802)。

摘　　要：针对现有无人驾驶方程式赛车目标检测算法运算量大、精度低等问题,提出方程式赛车目标检测算法。在图像预处理阶段,裁剪图片中无正样本部分,图片被小幅随机缩放平移。在网络结构上,调整ShuffleNetv2的结构,加强对颜色、光照和边缘等浅层特征的关注,利用特征金字塔对输出特征进行融合处理,基于广义焦点损失优化损失函数,获取正样本的类别和位置信息。实验结果表明:在FSACOCO数据集,文中算法的平均精度达到97.9%,浮点运算量为1.14 GFLOPs,优于其他对比算法。To solve the issues of high computation and poor accuracy of existing unmanned formula rac‐ing target identification algorithms,Formula car target detection algorithm was presented.In the image pre-processing stage,the part without positive samples in the image was cut off,and the image was panned with a small random zoom.The network structure of ShuffleNetv2 was adjusted to improve the focus on shallow features such as color,lighting,and edges.The output features were fused by using a feature pyramids network(FPN),the loss function was optimized by using generalized focal loss(GFL),and the classification and position information of positive samples were acquired.Experimental results show that,in FSACOCO data set,the average accuracy of this algorithm is 97.9%,and the floating�point computation is 1.14 GFLOPs,which are better than other comparison algorithms.

关 键 词：ShuffleNetv2 FPN GFL 锥桶检测 轻量化 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP183[自动化与计算机技术—计算机科学与技术]