面向软体机械臂抓取的单目深度估计方法

Monocular Depth Estimation Method for Soft Manipulator Grasping

软体机械臂在非结构化环境中执行物体抓取任务时,具有柔顺性和安全性的优势。但由于传统的立体视觉系统难以安装在软体机械臂上,软体机械臂获取目标物体的三维位置仍存在挑战。首先,建立了描述软体机械臂运动状态和目标物体位置的协作坐标系模型。随后,提出了一种改进运动恢复结构的单目深度估计方法:①针对深度估计的实时性要求,提出采用主成分分析-尺度不变特征变换(principal component analysis-scale invariant feature transform,PCA-SIFT)方法对特征描述符降维;②将机械臂末端的移动距离作为约束恢复深度估计的真实尺度。最后,通过实验验证了所提出的协作坐标系和深度估计方法的有效性。实验结果表明,所提深度估计方法在多种光照和遮挡环境中,特征匹配时间缩短了16.02%,平均误差控制在5 mm以内。

Soft manipulators have advantages of flexibility and safety when executing object grasping tasks in unstructured environments.However,traditional stereo vision systems are challenging to install on soft manipulators,making it difficult to obtain the three-dimensional position of target objects.Firstly,a collaborative coordinate system model describing the motion state of the soft manipulator and the position of the target object is established.Secondly,an improved structure from motion monocular depth estimation method is proposed:①Addressing the real-time requirement of depth estimation,principal component analysis-scale invariant feature transform(PCA-SIFT)is employed to reduce the dimension of feature descriptors;②The moving distance of the manipulator’s end effector is utilized as a constraint to recover the depth estimation at the real scale.Finally,experiments are conducted to validate the effectiveness of the proposed collaborative coordinate system and depth estimation method.The experimental results demonstrate that the proposed depth estimation method reduced feature matching time by 16.02%in various lighting and occlusion environments,with an average error controlled within 5mm.