4DRaL: Bridging 4D Radar with LiDAR for Place Recognition Using Knowledge Distillation
Ningyuan Huang, Zhiheng Li, Zheng Fang
AI summary
Problem
4D radar offers all-weather place recognition but suffers from inherent noise and sparsity that degrade performance, while existing cross-modal methods struggle with modality differences and training inefficiency.
Approach
The framework uses a high-performance LiDAR-to-LiDAR model as a teacher to guide a 4D radar student model through three distillation modules: local image enhancement, feature distribution alignment, and margin-based response distillation.
Key results
- State-of-the-art performance on NTU4DRadLM and SJTU4D benchmarks
- First knowledge distillation framework for 4D radar-to-LiDAR place recognition
- Superior robustness in adverse weather (rain, snow, fog) compared to LiDAR
- Balanced performance and computational efficiency via streamlined design
Why it matters
Enables reliable global localization for robotics and autonomous systems in harsh environments where optical sensors fail.
Abstract
Place recognition is crucial for loop closure detec- tion and global localization in robotics. Although mainstream algorithms typically rely on cameras and LiDAR, these sensors are susceptible to adverse weather conditions. Fortunately, the recently developed 4D millimeter-wave radar (4D radar) offers a promising solution for all-weather place recognition. However, the inherent noise and sparsity in 4D radar data significantly limit its performance. Thus, in this paper, we propose a novel framework called 4DRaL that leverages knowledge distillation (KD) to enhance the place recognition performance of 4D radar. Its core is to adopt a high-performance LiDAR-to-LiDAR (L2L) place recognition model as a teacher to guide the training of a 4D radar-to-4D radar (R2R) place recognition model. 4DRaL comprises three key KD modules: a local image enhancement module to handle the sparsity of raw 4D radar points, a feature distribution distillation module that ensures the student model generates more discriminative features, and a response distilla- tion module to maintain consistency in feature space between the teacher and student models. More importantly, 4DRaL can also be trained for 4D radar-to-LiDAR (R2L) place recognition through different module configurations. Experimental results prove that 4DRaL achieves state-of-the-art performance in both R2R and R2L tasks regardless of normal or adverse weather.