摘要
目前,工业机器人识别可抓取物品大多是先通过图像传感器收集作业场景信息,然后通过粒子滤波或条件随机场等各类相关算法提取可抓取物品的像素块特征来进行的。但是,这些可抓取物品的识别方法都存在着在同一像素块内部不同类别像素有误差,只考虑邻近区域、而不考虑全局信息和结构信息等问题或缺点。为此,在引入基于像素点的全卷积网络(fully convolutional networks,FCN)的基础上,提出了基于FCN的改进模型进行可抓取物品识别,其优势在于该模型经过学习能够预测各个像素所属物品类别的概率,并且能将结果恢复成为背景与前景分割的图像,从而识别作业场景中可抓取物品的位置与类别。由于FCN模型不限制输入、输出图像的尺寸大小,所以它克服了传统卷积深度学习模型的缺点,同时也考虑了全局信息与结构信息。以康奈尔抓取数据集(cornell grasping dataset,CGD)作为实验样本对提出的改进模型进行验证。实验结果表明:改进后的全卷积深度学习模型的正确率较全卷积深度学习模型提高了6.2%,同时该方法也可用于其他分割前景的感知任务。
Most of the traditional industrial robots collected environmental information information through the image sensor,and used the particle filter or the conditional random field and other algorithms to extract the feature of the pixel blocks for the grabbed object detection. These methodswas lack of the consideration of global information and structural information,deviation existed in the feature of the pixel block. In this research,taking Cornell Grasping Data set as experimental sample and applying fully convolutional networks based on pixels,we proposed an improvement model of fully convolutional networks for graspable object detection. The advantage of the model is predicting the category probability of each pixel by a learning way and spliting the output image into background and foreground. And then it gets the position and category of the graspable object. Since the full convolutional model did not limit the size of the input and output images,it overcame the shortcomings of the convolutional deep learning model,while the full convolutional model considered the global information and structural information. Experiments showed that our improvement fully convolutional deep learning model archived 6. 2% higher than the fully convolutional deep learning model（ fcn-8 x）.The proposed method can be used in other foreground segmentation perception tasks.
作者
皮思远
唐洪
肖南峰
PI Siyuan, TANG Hong, XIAO Nanfeng(School of Computer Science ＆ Engineering, South China University of Technology, Guangzhou 510006, Chin)
出处
《重庆理工大学学报（自然科学）》
CAS
北大核心
2018年第2期166-173,共8页
Journal of Chongqing University of Technology：Natural Science
基金
国家自然科学基金资助项目(61573145)
广东省公益研究与能力建设专项资金资助项目(2014B010104001)
广东省省自然科学基金资助项目(2015A030308018)
关键词
深度学习模型
全卷积网络
物品识别
工业机器人
deep learning model
fully convolutional network
object detection
industrial robot
