改进型Faster R⁃CNN的AGV导航图案目标检测算法  

Improved Faster R⁃CNN for target detection algorithm of AGV navigation pattern

作　　者：张洪涛[1] 田星星 周意入 秦宇 ZHANG Hongtao;TIAN Xingxing;ZHOU Yiru;QIN Yu(School of Electrical and Electronic Engineering,Hubei University of Technology,Wuhan 430068,China)

机构地区：[1]湖北工业大学电气与电子工程学院,湖北武汉430068

出　　处：《现代电子技术》2022年第13期51-56,共6页Modern Electronics Technique

基　　金：武汉市科技局“十城千辆新动力汽车计划”项目(2013011801010600)。

摘　　要：AGV视觉导航定位技术目前大多是在AGV的预设轨道上铺设含目标对象的导航图案,在拍摄到导航图案后先利用目标检测算法检测其目标区域,然后用角点检测算法提取目标区域的参考角点,最后利用参考角点和工业相机的焦距等参数的几何关系计算出AGV的位姿。文中在目标检测算法中经典的Faster R⁃CNN网络模型基础上加以改进,在多层次的feature map上生成候选框且用两个3×3卷积核分别进行卷积运算,从而直接进行分类和回归。仿真测试结果显示:相比Faster R⁃CNN,改进型Faster R⁃CNN检测所设计导航图案的mAP值提高了0.032,FPS值提高了31。因此证明改进型Faster R⁃CNN的精确度和速度均提高了,应用到AGV视觉导航定位技术中可进一步提高该技术的精确度和速度。At present,most of the AGVs(automatic guided vehicles)involve laying navigation patterns with target objects on the preset track of AGV.After taking the photographs of the navigation pattern,the target area is detected by the target detection algorithm,and then the reference angular point of the target area is extracted by the angular point detection algorithm.The pose of AGV is calculated by the geometrical relationship between the reference angular point and the focal length of industrial camera.The classical Faster R⁃CNN network model in the target detection algorithm is improved.The candidate boxes are generated on multi⁃level feature map,and two 3×3 convolutional kernels are used for convolution operation respectively,so as to directly classify and regress the generated anchor boxes.The simulation results show that,when the improved Faster R⁃CNN is used to detect the designed navigation pattern,the mAP(mean average precision)value is increased by 0.032 and FPS(frames per second)value is increased by 31 in comparison with those of the Faster R⁃CNN.Therefore,it is proved that the accuracy and velocity of the improved Faster R⁃CNN are both improved.When it is applied to AGV visual navigation and positioning technology,the accuracy and velocity of the technology can be improved furthermore.

关 键 词：目标检测算法 AGV导航图案 改进型Faster R⁃CNN 视觉导航 角点提取 AGV位姿计算 候选框生成 卷积运算 

分 类 号：TN911.73-34[电子电信—通信与信息系统] TP120.30[电子电信—信息与通信工程]