An Improved Point-to-Feature Recognition Algorithm for 3D Vision Detection

Vision-detection-based grasping is one of the research hotspots in the field of automated production. As the grasping scenes become more and more diversified, 3D images are increasingly chosen as the input images for object recognition in complex recognition scenes because they can describe the morphology and pose information of the scene target objects more effectively. With object recognition and pose estimation in 3D vision as the core, this paper proposes an improved pose estimation algorithm based on the PPF feature voting principle for the problems of low recognition rate and poor real-time performance in vision detection systems. The algorithm firstly performs preprocessing measures such as voxel downsampling and normal vector calculation on the original point cloud to optimize the point cloud quality and reduce the interference of irrelevant data. Secondly, an improved point cloud downsampling strategy is proposed in the point cloud preprocessing stage, which can better preserve the surface shape features of the point cloud and avoid introducing a large number of similar surface points. Finally, an improved measure of scene voting ball is proposed in the online recognition stage. The recognition and matching experiments on the public dataset show that the proposed algorithm has an average recognition rate improvement of at least 0.2%.