A Hybrid Data Association Method for Multi-Sensor SLAM

Abstract

Autonomous vehicles require simultaneous localization and mapping (SLAM). In order to create a more complete system, various sensors have been recently added, such as LiDAR, RGB-D cameras, inertial measurement units, and radar. A common issue when using multiple sensors is that the data obtained is inconsistent due to factors such as noise, offsets, or drift. This paper proposes a comprehensive hybrid data association framework. The framework integrates probabilistic modeling, graph-based optimization, and deep feature learning into a closed-loop system. Attention-based fusion modules and entropy-driven confidence gating are used to reliably match multi-sensor observations in a common latent space. Experimental validation through complex public and self-collected benchmarks has shown significant progress with the hybrid method. In terms of data association, a recall rate of 94.1% and an accuracy of 93.7% were achieved, with the average translational error reduced to 0.21 meters, outperforming both deep learning and traditional SLAM benchmarks. Real-time performance, memory usage: 1.2-1.5 GB, computation time per cycle: 48 ms. Ablation studies indicate that probabilistic reasoning, semantic encoding, and global graph optimization are the three components of the system, and they must work together to achieve robustness and mapping performance. The new model is not well-suited for areas with blur or severe occlusion, but it performs well in dynamic, crowded, and low-texture areas. This study sets new high standards for multimodal SLAM data association and provides practical support for the deployment of long-term, highly reliable robotic systems.