基于双目相机与改进YOLOv3算法的果园行人检测与定位  

Orchard Pedestrian Detection and Location Based on Binocular Camera and Improved YOLOv3 Algorithm

作　　者：景亮[1] 王瑞 刘慧[1] 沈跃[1] JING Liang;WANG Rui;LIU Hui;SHEN Yue(School of Electrical and Information Engineering, Jiangsu University, Zhenjiang 212013, China)

机构地区：[1]江苏大学电气信息工程学院,镇江212013

出　　处：《农业机械学报》2020年第9期34-39,25,共7页Transactions of the Chinese Society for Agricultural Machinery

基　　金：江苏省重点研发计划项目(BE2018372);江苏省自然科学基金项目(BK20181443);江苏高校青蓝工程项目、镇江市重点研发计划项目(NY2018001);江苏省三新工程项目(NJ2018-12)。

摘　　要：针对复杂果园环境中行人难以精确检测并定位的问题,提出了一种双目相机结合改进YOLOv3目标检测算法的行人障碍物检测和定位方法。该方法采用ZED双目相机采集左右视图,通过视差原理获取图像像素点的距离信息;将双目相机一侧的RGB图像作为用树形特征融合模块改进的YOLOv3算法的输入,得到行人障碍物在图像中的位置信息,结合双目相机获得的像素位置信息计算出相对于相机的三维坐标。用卡耐基梅隆大学国家机器人工程中心开放的果园行人检测数据集测试改进的YOLOv3算法,结果表明,准确率和召回率分别达到95.34%和91.52%,高于原模型的94.86%和90.19%,检测速度达到30.26 f/ms。行人检测与定位试验表明,行人障碍物的定位在深度距离方向平均相对误差为1.65%,最大相对误差为3.80%。该方法具有快速性和准确性,可以较好地实现果园环境中的行人检测与定位,为无人驾驶农机的避障决策提供依据。The accurate identification and location of obstacles in agriculturalenvironment is one of the most important technologies for intelligent agricultural machinery.Aiming at the problem that pedestrians are difficult to detect and locate accurately in the complex orchard environment,a method of pedestrian obstacle detection and location based on binocular camera and improved YOLOv3 target detection algorithm was proposed.In this method,the left and right views were collected by zed binocular camera,and the distance information of image pixels was calculated based on parallax principle.One side of the RGB image was used as the input of the improved YOLOv3 algorithm which by introduced the tree feature fusion module,and the position information of pedestrian obstacles in the image was obtained.And then the three-dimensional coordinates relative to the camera were calculated based on the pixel position information obtained by the binocular camera.Experiment carried on the open pedestrian detection dataset in orchard environment of the National Robotics Engineering Center of Carnegie Mellon University which contained different motion states(motion and static),different pose states(normal and unnormal)and different object scales(large,medium and small).Results showed that the average precision and recall rate of the improved YOLOv3 pedestrian detection model in agriculture reached 95.34%and 91.52%,respectively,which were higher than that of the original model(94.86%and 90.19%),and the detection speed was 30.26 f/ms.Meanwhile,the positioning accuracy of pedestrian obstacles was 1.65%in Z direction,and 3.80%in maximum.This method can locate pedestrian accurately and fast,providing reliable information for the obstacle avoidance of the unmanned agriculture machinery.

关 键 词：果园环境 行人检测定位 双目相机 YOLOv3 特征融合 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] S24[自动化与计算机技术—计算机科学与技术]