Dynamic Recalibration in LiDAR SLAM: Integrating AI and Geometric Methods with Real-Time Feedback Using INAF Fusion

TL;DR

This paper tackles LiDAR SLAM in GNSS-denied environments by introducing Inferred Attention Fusion (INAF), a real-time fusion module that recalibrates the contribution of AI-based and geometric odometry using environmental feedback. It integrates three odometry pathways (ICP-QE, DL-PE, and a fused variant) within a unified framework and optimizes the trajectory with loop closure and graph optimization, evaluated on KITTI. The key contributions are the INAF mechanism, the joint training of fusion networks, and a comprehensive performance evaluation showing improved translation and rotation accuracy over traditional and AI-based SLAM baselines, plus adaptive modality weighting under diverse driving scenarios. This work advances robust autonomous navigation by enabling dynamic, data-dependent fusion that remains accurate through GNSS-denied operation and dynamic environments.

Abstract

This paper presents a novel fusion technique for LiDAR Simultaneous Localization and Mapping (SLAM), aimed at improving localization and 3D mapping using LiDAR sensor. Our approach centers on the Inferred Attention Fusion (INAF) module, which integrates AI with geometric odometry. Utilizing the KITTI dataset's LiDAR data, INAF dynamically adjusts attention weights based on environmental feedback, enhancing the system's adaptability and measurement accuracy. This method advances the precision of both localization and 3D mapping, demonstrating the potential of our fusion technique to enhance autonomous navigation systems in complex scenarios.

Figures (9)

• Figure 1: Comparison of network adaptation strategies: General Attention-Based Networks vs. Inferred Attention (Ours). This figure illustrates how each network type adapts to new data. General attention networks use static attention weights that are determined during the training phase and do not change in response to new environmental data. In contrast, our Inferred Attention model dynamically adjusts its attention weights in real-time, utilizing feedback from environmental changes to enhance adaptation and improve accuracy in mapping and localization tasks.
• Figure 2: Pipelines illustrating three configurations: two single-branch setups (ICP-QE and DL-PE) and one fused setup combining both ICP-QE and DL-PE branches, followed by the Odometry Prediction Block and the Loop Closure and Optimization Block.
• Figure 3: a) Ensemble Learning vs. b) Network Fusion. In ensemble learning, multiple networks are trained independently and remain frozen during combination. In contrast, network fusion involves training multiple networks together, allowing for integrated learning and mutual adaptation.
• Figure 4: Adaptation of networks to new data across different methods: (a) Incremental learning, where the network is updated with new data over time; (b) Attention-based networks, which utilize attention mechanisms to focus on relevant parts of the input; and (c) Inferred Attention (Ours), where network parameters are continuously adapted based on feedback from the environment.
• Figure 5: Overview of the three phases in our network under INAF: (a) Training phase, where the network learns from data; (b) Prediction phase, where the trained network is frozen and used for inference, and (c) Prediction phase with Inferred Attention Fusion, where some network parameters continue to adapt based on feedback from the environment.