【摘要】 作为国内外综合性最大的机器人赛事之一,RoboCup竞赛吸引了越来越多的高校学生参加。对于以足球比赛为背景的Kidsize组,如何准确地通过图像检测足球和球门成为了比赛的关键性任务之一。通过深度学习的方法,引入深度卷积神经网络YOLOv3-tiny构建机器人的视觉识别系统。针对YOLOv3-tiny识别精度不高的问题,提出在进行检测前先对草坪颜色进行归一化,同时基于二阶梯度算子对图像进行锐化处理,以增强图像中的语义信息。实验结果表明,相较于传统方法,所构建的基于深度学习的视觉识别系统的检测效果更好,对不同场地和光照的适应能力更强,在以Nvidia Jetson TX2为代表的边缘深度学习设备上有着不错的推理速度,能够满足参加比赛的要求。更多还原

【Abstract】 As one of the most comprehensive robotics events all over the world,the RoboCup competition has attracted more and more scholars to participate.For the Kidsize group with the football game as the background,how to accurately detect the football and goal through the image has become one of the key tasks of the game.In this paper,a deep convolutional neural network YOLOv3-tiny is applied to build a robot’s visual recognition system.Aiming at the problem that YOLOv3-tiny recognition accuracy for small objects is not high,it is proposed to normalize the lawn color before detection,and at the same time sharpen the image based on the two-step operator to enhance the semantic information in the image.更多还原