4DRaL: Bridging 4D Radar with LiDAR for Place Recognition using Knowledge Distillation

Abstract

Place recognition is crucial for loop closure detection 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 distillation 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.

Figures (6)

• Figure 1: (a) Comparison of runtime and performance across several methods on SJTU4D dataset 4DRadarDataset. (b) The framework of 4D radar place recognition based on knowledge distillation. The query scan is encoded into a descriptor, and the database is searched for the closest matching descriptors. The R2R (radar-to-radar) and L2L (LiDAR-to-LiDAR) databases consist of 4D radar descriptors and LiDAR descriptors, respectively.
• Figure 2: Comparison of the two paradigms of R2L (radar-to-LiDAR). The joint training paradigm employs one or more encoders to directly align the features of LiDAR and 4D radar, while the knowledge distillation paradigm leverages a teacher model to guide the feature alignment of two modalities.
• Figure 3: Overview of 4DRaL. Given LiDAR and 4D radar point clouds of the same scene, we generate BEV images. The teacher model extracts features from LiDAR BEV to guide the student model, which consists of R2R and R2L. R2R applies LIE for image enhancement, FDD for feature learning, and RD for refining output relationships. R2L enhances radar data with LIE and aligns LiDAR features via RD. The teacher model is used only during training.
• Figure 4: The influence of margin on Response Distillation (RD) module.
• Figure 5: Visualization of Local Image Enhancement. In the area highlighted by the blue ellipse, the enhanced 4D Radar BEV image presents a clearer outline compared to the original image.