Feature Extraction Research In 3D Point Cloud Analysis

As a key sensor information for autonomous driving tasks,3D point cloud data provides a rich description of three-dimensional scenes.Point cloud analysis,as the foundation of 3D vision tasks,still faces issues of complex network structure and insufficient efficiency.Although recent compact Multi-Layer Perceptron(MLP)structure networks have shown potential,there remain two critical issues:firstly,the limited expressive ability of standard MLPs affects the effectiveness of local feature aggregation;secondly,the high computational complexity and lack of attention to surface information in existing point cloud local feature aggregation methods constrain network performance and efficiency.To address these issues,the main contributions and innovations of this paper are as follows:We propose a vector-based local feature aggregation method to address the limited expressive ability of standard MLP aggregation,in which scalar representation of neighbor importance is insufficient.In this method,we treat the scalar components of feature vectors as base vectors on coordinate axes and introduce a high-dimensional vector transformation to improve neighbor relationship representation,converting the neighbor aggregation problem into a problem of vector summation with co-directional promotion and counter-directional suppression.Simultaneously,we point out the importance of independent quantities in network optimization and transformation degrees of freedom,decomposing vector transformation into scaling and rotation transformations,and introducing independent rotation angles to construct rotation matrices,thus reducing network optimization difficulty and parameter quantity.The PointVector model built upon this method achieves optimal performance on major benchmark datasets while significantly reducing parameter quantity.We propose a graph-based sampling method to address the high computational complexity of existing Euclidean distance-based sampling methods and the destruction of point cloud surface structure information due to repeated establishment of neighborhood relationships during up-sampling and down-sampling.By fitting local surfaces with planes,we introduce a graph structure and establish point cloud relationships using local neighborhood connections,representing surface distances using graph node distances.Furthermore,we introduce a voxelization method to accelerate the graph construction process and design a graph structure update algorithm for down-sampling to preserve surface structure information,avoiding repetitive establishment of neighbor relationships and redundant neighborhood partitioning calculations.The enhanced PointVector model using this sampling method achieves better performance on major benchmark datasets,with unchanged parameter quantity and improved inference speed.The methods proposed in this study have been validated through detailed ablation experiments,providing valuable insights for point cloud analysis.