Bio-Inspired Intelligent LiDAR Odometry via Fast and Slow Cognitive Processing

Continuous and reliable state estimation and mapping are vital prerequisites for intelligent vehicle systems. While modern LiDAR odometry systems have achieved remarkable success in balancing accuracy and real-time performance, they often rely on rigid processing strategies that lack the adaptability needed to handle varying task complexities in real-world scenarios. In this paper, we propose an intelligent LiDAR odometry framework via Fast and Slow Cognitive Processing to address this challenge. Inspired by the human cognitive model of “Fast and Slow Thinking”, our method introduces a dual-path estimation framework that dynamically switches between a Fast-Path and a Slow-Path based on task complexity. The Fast-Path is designed for rapid state estimation in relatively stable environments, providing real-time performance through lightweight forward propagation and coarse matching. In contrast, the Slow-Path leverages a robust Iterated Extended Kalman Filter (R-IEKF) for high-precision updates in more complex or uncertain scenarios. A switching mechanism evaluates task complexity, considering residual errors, velocity changes, and point cloud structure, and determines when to switch between the two paths. Additionally, we propose a novel Similarity Pruning mechanism that curbs local map growth and further accelerates the Fast-Path. This approach ensures that important features are prioritized, while redundant data is minimized, improving both computational efficiency and estimation accuracy. Extensive experiments demonstrate that the proposed method outperforms conventional LiDAR odometry techniques, providing a more flexible and efficient solution for real-time state estimation in a wide range of real-world applications.