基于空间信息聚合的遮挡目标抓取位姿检测

Occluded target grasping detection method based on spatial information aggregation

针对机器人依靠视觉抓取时对遮挡目标抓取位姿检测准确率低的问题,提出基于空间信息聚合的遮挡目标抓取位姿检测方法。遮挡导致目标在相机视野中的本征特征改变,影响目标位置信息与形状结构特征。首先,使用坐标卷积代替传统卷积的方式进行特征提取,在输入特征图后新增坐标通道来提升网络对位置信息感知能力;其次,设计空间信息聚合模块,其采用并行结构增大局部感受野并沿空间方向对通道进行编码获取多尺度空间信息,再通过非线性拟合方式将信息聚合,使模型更好理解目标结构和形状;最后,抓取位姿检测网络输出抓取的质量、角度和宽度,并计算最佳抓取位置以建立最优抓取矩形框。在Cornell Grasping数据集、自建遮挡数据集、Jacquard数据集验证,检测准确率分别达到98.9%,94.7%,96.0%,在实验平台对目标的100次真实抓取实验中,成功率为93%。所提方法在三个数据集上均取得了最高检测准确率,且在实际场景中检测效果更优。

Aiming at the problem of low accuracy of occlusion target grasping position detection when the robot relies on vision grasping,we proposed an occlusion target grasping position detection method based on spatial information aggregation.Occlusion led to the change of the target's intrinsic features in the camera's field of view,which affected the target's positional information and shape-structural features.First,coordinate convolution was used instead of traditional convolution for feature extraction,and a new coordinate channel was added after the input feature map to improve the network's ability to perceive position information.Second,the spatial information aggregation module was designed,which adopted a parallel structure to increase the local sensing field and encoded the channels along the spatial direction to obtain multi-scale spatial information,and then aggregated the information through nonlinear fitting to make the model better understand the target structure and shape.Finally,the grasping position detection network outputted the grasping mass,angle and width,and calculated the optimal grasping position to establish the optimal grasping rectangular box.Validated on the Cornell Grasping dataset,the self-constructed occlusion dataset,and the Jacquard dataset,the detection accuracies reach 98.9%,94.7%,and 96.0%,respectively,and the success rate is 93%in 100 real grasping experiments on the target in the experimental platform.The proposed method achieves the highest detection accuracy on all three datasets,and the detection effect is better in real scenes.