An improved geometric priors-based indoor 3D object detection in point clouds for mobile robots

Abstract

3D object detection is an essential and highly challenging task for unmanned autonomous systems operating in indoor scenes. Current mainstream 3D detection approaches rely on the direct encoding of point cloud coordinates, while neglecting the underlying geometric priors, thereby limiting the expressiveness of learned features. To deal with these problems, an efficient indoor 3D object detection model (Geo3D) is proposed in this paper, which exploits geometric priors to refine the feature learning pipeline. Specifically, during the initial stage of feature extraction, a geometric invariant embedding module is designed to compute geometric invariants that are robust to translation and rotation. These invariants provide structural priors of the point cloud distributions, and are embedded into voxel-wise feature representations. Subsequently, a hierarchical encoder is presented to capture multi-scale contextual features, augmented with geometric-aware refinement layers, where geometric information serves as conditional priors to guide fine-grained feature learning. Then, a synergistic attention calibration module is integrated into the generative decoder, which utilizes staged attention mechanisms to adaptively align high-layer semantic features with low-layer geometric details. Finally, extensive experiments are conducted on the SUN RGB-D and ScanNet V2 datasets, and the results demonstrate that the proposed Geo3D model achieves superior performance compared to existing point cloud–based detection methods.