A Generative Hierarchical Optimization Framework for LiDAR Odometry Using Conditional Diffusion Models

LiDAR odometry serves as a cornerstone for environmental perception and spatial localization in emerging virtual-real integrated human-machine interaction (HMI) systems. While recent deep learning-based discriminative LiDAR odometry methods achieve reasonable accuracy, they predominantly optimize the data term of the Maximum A Posteriori (MAP) formulation while neglecting explicit prior modeling, thereby limiting their performance upper bound. This paper presents PWDLO, a generative LiDAR odometry framework that establishes the first hierarchical optimization architecture integrating explicit data-prior co-modeling through diffusion probabilistic principles. Our three-tier Pyramid-Warping-Diffusion (PWD) mechanism enables progressive pose refinement from coarse to fine granularity. Specifically, a pose refinement layer based on a conditional denoising diffusion model is proposed to inject the priors within pose generation process. To restrain the generation diversity and achieve fine-grained denoising, two pose-related features are leveraged as conditions in our diffusion model, and iteratively update a noisy pose residual through a recurrent unit that performs lookups on the conditions. Experiments on the KITTI odometry dataset demonstrate that, compared to state-of-the-art deep LiDAR odometry methods that directly use raw 3D point clouds, PWDLO reduces the average rotational and translational RMSE by 4.1% and 22.8%, respectively. Additionally, PWDLO outperforms LOAM on most sequences. To our knowledge, this is the first LiDAR odometry network based on diffusion probabilistic model. It demonstrates desirable characteristics that are challenging for discriminative models to achieve, such as progressive inference, uncertainty estimation and variation in LiDAR sensor performance.