基于元学习和关键点的实时抓取检测算法

Real-time Grasp Detection Algorithm Based on Meta-learning and Key Points

目前大多数深度神经网络在机器人抓取检测任务上取得了较好的检测效果,但对于未见过的物体抓取预测具有挑战性,且平衡抓取准确率和实时性也是一个难题。针对上述问题,该文提出一种基于元学习和关键点的实时抓取检测算法,将训练过程分为基础学习和元学习2个阶段。基于CenterNet思想设计网络结构,引入元学习器增强模型对未知知识的学习能力。首先采用旋转椭圆高斯热力图预处理标签;然后基于GhostNet中的模块和NAM注意力机制设计轻量化的特征提取网络;为提升不同尺寸物体检测效果,设计多尺度的特征提取和特征融合模块。网络参数量和计算量分别低至4.08 M和3825 M,保证了实时性。在Cornell数据集上按照图像分割和对象分割进行实验,分别达到99.3%和99.0%的准确率。

Most deep neural networks have shown satisfactory performance in robot grasping detection tasks,but predicting the grasping of unseen objects remains a challenge,and striking a balance between grasping accuracy and real-time performance is also problematic.To tackle these issues,a real-time grasping detection algorithm is proposed,which is based on meta-learning and key points.The learning process is divided into a basic training phase and a meta-learning phase.The network structure is designed based on the CenterNet idea,and a meta-learner is introduced to enable the model to adapt to unknown knowledge.Firstly,the improved rotating ellipse Gaussian heatmap is adopted in the preprocessing stage.Secondly,the lightweight feature extraction network uses the Ghost Bottleneck embedded with the NAM attention.Finally,a multi-scale feature extraction and fusion module is designed to improve the detection effect of objects with different sizes.The parameters and FLOPs are kept as low as 4.08 M and 3825 M,respectively,which can ensure real-time performance.The algorithm achieves high accuracies of 99.3%and 99.0%on the Cornell dataset for image-wise split and object-wise split,respectively.