Pose Estimation Of Semantic Components Based On Deep Learning And Its Application

The involvement of robotic arms in the manufacturing process is indispensable,and in recent years,the precision requirements of their operations have been constantly increasing.To address the research needs for object-level operations,we have analyzed the impact of semantic object parts on the 6D pose estimation and further improved the corresponding neural network models.In practice,we have employed a series of strategies to enhance the precision of the robotic arms’ object-level operations.The main work of this paper can be roughly divided into three parts:In the analysis of pose estimation accuracy based on object part segmentation,we started with the assumption that the effectiveness of 6D pose estimation for an object may depend on its own topological structure.With the help of current leading benchmark network models,we conducted pose estimation analyses on both the object as a whole and its individual parts.To this end,we constructed a dedicated RGB-D dataset for part-level object 6D pose estimation,consisting of 100,000 samples,each with 6 instances and a total of 12 parts,along with part-level pose annotations.Based on this dataset,we performed experiments on both whole object and part-level pose estimation,and analyzed the impact of part relationships on the accuracy of whole object pose estimation.Using these conclusions,we improved the benchmark network model,resulting in a 2.1% increase in pose estimation accuracy,thus providing initial evidence supporting our conclusions.In the section dedicated to 6D pose estimation based on part attention mechanism,we have proposed a more effective method called SURF-FPS for selecting 3D keypoints of object models.This method takes into account both the weak texture information on the object model surface and the geometric constraints of the model vertices.Following this,we have introduced the CPFNet network model,which includes a point cloud feature encoding module that efficiently enhances local context features of the target object,a module that fuses high-level features from images and depth,a semantic part-weighted composite instance pose module,and a corresponding set of loss functions.We have demonstrated the outstanding pose estimation performance of the CPFNet model through comparative experiments on a dedicated dataset,achieving a comprehensive instance average pose estimation accuracy of 81.0%.This represents a 4.7% improvement compared to the baseline model,and the effectiveness of each module has been verified through ablation experiments.In the section of ”Precise Grasping of Mechanical Arm Based on Multi-View Semantic Parts Pose Estimation”,we proposed a optimization strategy for pose estimation based on the motion trajectory of objects on the conveyor belt,and achieved high-precision pose estimation of part-level positioning in real conveyor belt scenes using this strategy and the CPFNet model.We also designed an automated pose matrix transformation strategy based on object-part relationships,which can automatically determine the target part that is best suited for grasping while transforming the overall pose into the pose of the target part.Based on this,we built a part-level object manipulation system mainly using industrial robotic arms,and through sufficient part grasping experiments,we demonstrated the effectiveness of the CPFNet model and the aforementioned strategies.The accurate grasping rate of this system is 96%.This work highlights how the combination of motion trajectory and semantic part information improve the precision of grasping,and provides new insights for related research in the field of robotics.