Trans-LGS: Transformer-Based Local Graph Structure for 3D Object Detection in Autonomous Vehicles

As intelligent and interconnected vehicles continue to progress, the importance of 3D LiDAR sensors in traffic perception is increasing. Accurate 3D object detection in point clouds is crucial for autonomous vehicles. Current methods focus on local feature extraction, disregarding vital feature interactions, leading to inadequate semantic information and suboptimal detection performance. We present Trans-LGS, a Transformer-based local graph structure. It employs a Graph Neural Network (GNN) and self-attention to learn semantic feature information by aggregating neighboring data and highlighting significant relationships relative to center point features within the local graph. Trans-LGS consists of four modules: Graph Feature Abstraction (GFA), Semantic Features Extraction (SEFE), Spatial Features Extraction (SPEF), and Local and Global Features Aggregation (LGFA). GFA establishes a local graph structure by selecting a center point feature and utilizing k-nearest neighbors. SEFE extracts neighboring features to learn crucial object points. SPEF calculates relative positions and distances to capture latent position information. LGFA combines low-dimensional global and high-dimensional local features. Comprehensive trials on the KITTI and DIARV2X-I datasets exhibit significant enhancements in established techniques (SECOND and PointPillars). Notably, 3D average precision (AP) increases by 6.35% and 6.72% on the KITTI test dataset. Introducing Trans-LGS improves and stabilizes performance for medium and long distances. Crucially, Trans-LGS integration minimally impacts real-time performance and model parameters.