在工业生产中，待抓取物体往往具有种类众多、摆放位置杂乱、形状不规则等特点，使得难以准确获取物体抓取位姿。针对以上问题，提出一种基于深度学习的两阶段抓取位姿估计方法。第1阶段，提出一种基于YOLOv4(you only look once version4)改进的轻量级旋转目标检测算法，提高目标的检测速度和检测精度。首先，使用轻量化网络GhostNet和深度可分离卷积对原始网络进行重构，降低整个模型参数。然后，在颈部网络中增加自适应空间特征融合结构和无参注意力模块，提高对感兴趣区域的定位精度;最后，使用近似倾斜交并比(skew intersection over union，SkewIoU)损失解决角度的周期性问题。第2阶段，制作与原始图片尺寸一样的掩膜提取感兴趣区域;同时，提出一种改进的DeepLabV3+算法，用以检测感兴趣区域中物体的抓取位姿。实验结果表明，改进后的YOLOv4网络检测精度达到92。5％，改进的DeepLabV3+算法在Cornell抓取数据集上的图像拆分和对象拆分精度分别达到94。6％，92。4％，且能准确检测出物体的抓取位姿。
Object grasp pose detection based on the region of interest
In industrial production,the objects to be grasped often have the characteristics of varions types,messy placements,irregular shapes,etc.,which make it difficult to accurately obtain the grasping pose of the object.In view of the above problems,this paper proposes a two-stage grasp pose estimation method based on deep learning.In the first stage,a lightweight rotating target detection algorithm based on improved you only look once version4(YOLOv4)is proposed to enhance the detection speed and improve detection accuracy of targets.Firstly,the lightweight network GhostNet and deep separable convolution are used to reconstruct the original network to reduce the parameters of the entire model.Then,the adaptive spatial feature fusion structure and the non-reference attention module are added to the neck network to improve the positioning accuracy of the region of interest.Finally,the approximate skew intersection over union(SkewIoU)loss is used to solve the periodic problem of the angle.In the second stage,a mask extraction region of interest is made with the same size as the original picture.At the same time,an improved DeepLabV3+algorithm is proposed to detect the grasping pose of objects in the area of interest.Experimental results show that the detection accuracy of the improved YOLOv4 network reaches 92.5％,and the improved DeepLabV3+algorithm achieves 94.6％and 92.4％of the image splitting and object splitting accuracy on the Cornell capture dataset,respectively,and can accurately detect the grasping pose of objects.
deep learningmaskregion of interestlightweight networkpose detection
NETL
NSTL
万方数据
