Research On Object Pose Estimation Method For Solid Waste Sorting

With the development of industrial intelligence,the recycling method for solid waste has gradually shifted from manual sorting to efficient and safe intelligent sorting.Among them,detecting the category and posture of the sorted materials is the key foundation for achieving target grasping.In order to reduce the limitations of hardware computing power and storage capacity on the application of target pose estimation in industrial production,reduce the number of network model parameters for target detection,and meet real-time classification requirements.At the same time,the mutual occlusion between objects in sorting scenarios poses difficulties in estimating the spatial pose of objects.It is of great significance to study how to reduce the loss of feature information and improve the accuracy of pose estimation models.The specific research content of this article is as follows:1)Research on lightweight object detection methods.Due to the large number of parameters in the existing YOLO v4 backbone network,it is difficult to deploy in embedded devices with limited resources.In order to reduce the parameters and computational complexity of deep neural networks,a feature extraction network for object detection is constructed using structures such as deep separable convolution and inverted residual blocks.In order to ensure model accuracy,the feature pyramid pooling module is improved by using hollow convolution and SE channel attention modules,Construct a YOLO v4-MN lightweight object detection model.Extract images from the YCB Video dataset and produce them as object detection datasets to test the performance of lightweight object detection networks.2)Research on pose estimation methods for fusion point cloud completion.When using point cloud data with category information for pose estimation,due to the occlusion of the target object,some surface points cannot be observed,and the point cloud data lacks local details and edge information of the object,which limits the comprehensive understanding of the pose estimation network of the object.This article proposes a pose estimation network model that integrates point cloud completion to estimate possible point cloud data in occluded areas,providing more complete surface structure information for pose estimation.The multi-scale point cloud completion module adopts an encoder decoder structure to achieve low to high precision feature extraction and fusion,recover missing point cloud data,and perform independent regression on rotation and translation in the pose estimation module.Experimental verification was conducted on the YCB Video dataset,and the experimental results showed that compared to the method of directly estimating the pose of incomplete point clouds,The fusion of point cloud completion methods can achieve higher pose estimation accuracy and effectively reduce the impact of incomplete spatial position information of the target object caused by occlusion.This article designs a lightweight feature extraction network in the target detection stage,which can achieve faster model operation efficiency and achieve real-time detection purposes;In the pose estimation stage,a pose estimation method that integrates multi-scale point cloud completion is proposed to address occlusion issues,which can effectively reduce the impact of occlusion on feature extraction.