DNOI-4DRO: Deep 4D Radar Odometry with Differentiable Neural-Optimization Iterations

TL;DR

DNOI-4DRO tackles robust 4D radar odometry by coupling a neural motion-field estimator with a differentiable Gauss-Newton-based optimizer, enabling end-to-end training. It introduces a dual-stream radar backbone with multi-scale geometric and clustering-based class-aware features, and an adaptive correlation-driven iteration module that refines pose over multiple steps. Experiments on VoD and Snail-Radar show state-of-the-art performance, with results competitive to LiDAR-based A-LOAM in short-range localization and substantial improvements over prior 4D radar methods. The work demonstrates the value of integrating geometric optimization into deep radar odometry and provides a public release of code.

Abstract

A novel learning-optimization-combined 4D radar odometry model, named DNOI-4DRO, is proposed in this paper. The proposed model seamlessly integrates traditional geometric optimization with end-to-end neural network training, leveraging an innovative differentiable neural-optimization iteration operator. In this framework, point-wise motion flow is first estimated using a neural network, followed by the construction of a cost function based on the relationship between point motion and pose in 3D space. The radar pose is then refined using Gauss-Newton updates. Additionally, we design a dual-stream 4D radar backbone that integrates multi-scale geometric features and clustering-based class-aware features to enhance the representation of sparse 4D radar point clouds. Extensive experiments on the VoD and Snail-Radar datasets demonstrate the superior performance of our model, which outperforms recent classical and learning-based approaches. Notably, our method even achieves results comparable to A-LOAM with mapping optimization using LiDAR point clouds as input. Our models and code will be publicly released.

Figures (5)

• Figure 1: Overview of our backbone. (1) The feature and context extractors encode the point and context features of the input point cloud, respectively. (2) The feature correlation module constructs an all-pair correlation volume by calculating the matrix dot product of two-point features. (3) In each iteration, the differentiable neural-optimization iteration operator uses the pose estimated in the previous iteration to look up correlation features from the correlation volume, which are then processed through a GRU gru to generate a point motion field. The point motion field is fed into a least-squares-based optimization layer, where the pose is updated based on geometric constraints. After multiple iterations, the network outputs the predicted pose.
• Figure 2: The structure of dual-stream radar feature extraction network.
• Figure 3: Illustration of the differentiable neural-optimization iteration operator, which predicts point motion flow revisions and maps them to pose updates through the AMBA layer.
• Figure 4: 3D and 2D trajectory results for VoD test sequences 19 and 22, and Snail-Radar test sequences st. Our method obtains the most accurate trajectory.
• Figure 5: Visualization of clustering results. 4D radar points are color-coded based on their assigned categories.