Abstract
LiDAR Odometry (LO) is a fundamental component of autonomous navigation, yet traditional and deep learning-based methods often struggle with dynamic noise and complex transformation distributions. This paper introduces DiffLO, a novel framework that achieves state-of-the-art accuracy by integrating semantic scene understanding with generative diffusion models. By distilling knowledge from a segmentation "teacher" and treating pose refinement as an iterative denoising process, DiffLO significantly outperforms existing deep learning methods and classic geometry-based algorithms like A-LOAM on the KITTI benchmark.
Introduction
LiDAR Odometry involves estimating a vehicle's motion by comparing consecutive 3D point cloud frames. While traditional geometry-based methods (e.g., ICP) are precise in static environments, they frequently fail in real-world conditions characterized by sparse data, dynamic objects like moving vehicles, or repetitive structures.
Existing deep learning approaches attempt to address these limitations but face two primary hurdles: semantic blindness and deterministic limits. Most models treat all points equally, failing to distinguish between stable landmarks (buildings) and dynamic noise (buses). Furthermore, standard models typically use simple Multi-Layer Perceptrons (MLPs) to regress a pose in a single pass, which often fails to capture the complex, noisy distribution of possible spatial transformations.
Methodology
DiffLO addresses these challenges through a two-pronged architectural strategy: the Semantic Perception Module and Diffusion-Based Refinement.
Semantic Perception Module (SPM)
To incorporate environmental awareness, DiffLO utilizes Knowledge Distillation. During the training phase, a pre-trained, heavy-duty 3D segmentation model acts as a "teacher," guiding the LO network to identify stable structures within the scene. Crucially, this teacher model is discarded during inference, allowing the system to remain lightweight and fast while retaining the ability to ignore dynamic noise.
Diffusion-Based Refinement
DiffLO treats pose estimation as a generative task rather than a simple regression. The model learns a reverse (denoising) process to iteratively "clean" a noisy pose estimate. This process is conditioned on geometric features, cost volumes, and distilled semantic features, ensuring the generative process remains grounded in the actual sensor data.
Experimental Results
The researchers evaluated DiffLO on the KITTI Odometry Dataset, using sequences 00-06 for training and 07-10 for testing. DiffLO sets a new benchmark for learning-based LiDAR odometry with the following key results:
- Superior Accuracy: DiffLO achieved a 10.9% improvement in translation accuracy and a 16.9% improvement in rotation accuracy over the previous state-of-the-art, EfficientLO.
- Competitive Performance: It outperformed A-LOAM, a widely used traditional method that includes back-end mapping optimization, across most evaluation sequences.
- Real-Time Efficiency: Despite the iterative nature of diffusion, the system achieves an inference time of 76.9 ms per frame, meeting the 100 ms requirement for real-time 10Hz LiDAR sensors.
Discussion and Future Work
DiffLO demonstrates that generative models are exceptionally well-suited for refining 6-DoF poses due to their iterative nature. The use of distillation provides a "free lunch," granting the model semantic awareness without the computational overhead of running a segmentation sub-network during deployment.
However, limitations remain. The model's semantic awareness is currently capped by the performance of the teacher model. Additionally, while real-time performance is achieved, scaling to higher-frequency sensors or lower-power hardware may require further optimization of the denoising steps. Future research will explore the generalization of this semantic distillation to diverse, non-urban environments without the need for extensive retraining.
Summary updated