LITE: Local Isometry-Invariant Topology Encoding for Efficient 3-D Sensing With 4-D Radar Point Clouds

In autonomous driving, 4D millimeter-wave (MMW) radar is increasingly used as a low-cost, all-weather onboard sensing solution for 3D spatial measurement of the surrounding environment. However, the inherent sparsity and noise in radar-generated point clouds introduce significant uncertainty in object-level spatial localization and measurement. To improve the accuracy and robustness of radar-based object detection, we propose Local Isometry-invariant Topology Encoding (LITE), a radar-exclusive perception architecture that explicitly encodes local isometry-invariant features. LITE incorporates a structured multi-stage pipeline optimized for sparse radar data measurements. It begins with early voxel-level feature aggregation to reduce input dimensionality while preserving spatial fidelity. The Adaptive Isometry-invariant Voxel Encoder (AIVE) extracts robust local geometric, velocity, and reflectivity distribution that remain consistent under rigid transformations. A Neighborhood Attention Module (NAM) transforms these handcrafted features into trainable, context-aware embeddings by modulating them with surrounding voxels. Without the need for auxiliary image data or dense manual annotations, LITE operates as a standalone radar-based sensing module capable of real-time 3D object detection. Evaluated on the TJ4DRadSet dataset, it achieves a 3D mean Average Precision (mAP) of 32.83% with a runtime of 56.3 FPS on an NVIDIA RTX 3090 Ti, demonstrating low-latency inference and reduced model size (with 4.848M parameters). These results demonstrate that explicitly encoding local isometry-invariant features enhances both spatial measurement accuracy and processing efficiency, supporting the use of 4D radar in compact, real-time sensing systems for autonomous applications.