Object Recognition and Spatial Relationship Determination in Point Clouds
Qingwei Song, Naoyuki Kubota · 2023
In smart living, scene interpretation is necessary. Determining the spatial relationships of each object helps the system understand people's needs. For monocular machine vision of 2D images, it is difficult to obtain accurate spatial relationships at some angles. The point cloud obtained by the depth camera has a natural advantage for such objects. However, the object recognition of 3D point cloud is computationally intensive, and the original image recognition accumulated data set is completely abandoned. This paper combines the advantages of image recognition and point cloud, uses the existing model of image recognition to identify objects, and accurately finds the point cloud of the object in the point cloud, and then determines the spatial relationship. This paper first uses the 2D object recognition algorithm to find out the approximate range of each object and then uses the GNG clustering algorithm to remove the background point cloud and identify the corresponding 3D object. This method can effectively improve the accuracy and efficiency of 3D object recognition. Combining the GroundingDINO and SAM methods, it is possible to recognize and segment the point clouds of various objects in 3D by simply inputting their category names. By finding the centroid of each point cloud cluster after recognition, the accurate spatial relationship between objects can be obtained.