融合Transformer的特征交互网络在部分点云配准中的应用OA

With the rapid advances of 3D sensing technology,point cloud data are widely used in such fields as autonomous driving,robotic navigation,and 3D reconstruction.Point cloud registration,as a core component of these applications,aligns point clouds from different perspectives or time points into a unified coordinated system by estimating rotation and translation parameters.Although registration technology evolves from traditional algorithms to deep learning methods with end-to-end advantages,there are still limitations in handling real-world challenges.Existing deep learning registration methods suffer insufficient robustness when facing severe occlusion,drastic viewpoint changes,and partial overlap.The root cause lies in the excessive reliance of mainstream models on explicit overlap mask estimation or hard extraction.This approach makes registration accuracy highly dependent on the precision of overlap prediction.Consequently,performance drops sharply if predictions fail in low-overlap or high-noise environments.Additionally,many methods lack effective early information exchange during feature extraction and thus fail to account for the distinct solution spaces of rotation and translation parameters in rigid transformations. To address these challenges,this paper introduces TFFNet,a Transformer-based end-to-end feature interaction network specifically designed to enhance accuracy and stability in partial point cloud registration.The architectural design of TFFNet acknowledges rotation and translation possess low correlation and occupy vastly different solution spaces.TFFNet ingenuously creates a dual-branch structure for feature extraction and regression.By modeling rotation and translation features in independent paths,the network improves the specificity of feature learning and lays a foundation for subsequent branch interaction. In the encoder stage,a sophisticated multi-level feature interaction mechanism compensates for the global modeling deficiencies of traditional CNNs.First,the network integrates the Source-reference Feature Fusion(SFF)module,which introduces a self-attention mechanism.This module employs sinusoidal position encoding to incorporate spatial information and uses Blocked Multi-head Self-attention to capture non-local dependencies within the input.This design enables the extraction of salient features containing rich global contextual information.Next,to achieve efficient alignment,the Cross-Attention Fusion(CFF)module treats source features as queries and reference features as keys and values.Dynamic information injection guides the source branch to perceive the structural details of the reference point cloud,markedly enhancing alignment capabilities.At the end of the feature extraction path,the Rotation-Translation Fusion(RTF)module concatenates and maps features from both branches following the pooling layer.This achieves coordination and unification of rotation and translation information in the global dimension,strengthening the overall perception of complex rigid transformations. Moreover,a dual-state loss function improves feature discriminability and geometric sensitivity.The loss function consists of two parts:the pose regression loss and the transformation sensitivity loss.The pose regression loss directly supervises the absolute accuracy of predicted quaternions and translation vectors.The transformation sensitivity loss minimizes the distance between positive pairs and maximizes the distance between negative pairs.This constraint forces the network to extract highly discriminative features across different transformation states,guiding it to focus autonomously on geometric structures critical to registration while reducing reliance on redundant information.