Automotive millimeter-wave (mmWave) radar is crucial for all-weather autonomous driving, making radar-camera fusion a highly promising perception solution. Conventional fusion frameworks rely on sparse radar point clouds generated via constant false alarm rate (CFAR) filtering, suffering from severe information loss. While recent raw-tensor-based approaches successfully bypass CFAR, they introduce prohibitive computational burdens and struggle with cross-modal spatial misalignments. To address these challenges, we propose RawRadarFusion, a highly efficient multimodal 3-D object detection paradigm that fundamentally shifts away from heavy tensor operations. First, we compress raw 4-D radar tensors into a lightweight motion saliency index (MSI) pseudoimage. This velocity-weighted representation preserves essential macrokinematics and elevation cues while drastically reducing data dimensionality. Second, we introduce the deformable columnwise cross-modal attention (DCCA) module, which leverages geometrically constrained learnable offsets to achieve precise spatial alignment between radar and camera features, effectively mitigating calibration drifts without the overhead of global attention. Extensive experiments on the K-Radar dataset demonstrate that RawRadarFusion not only achieves highly competitive accuracy against dense-tensor state-of-the-art models but also achieves an accuracy–efficiency tradeoff while exhibiting strong robustness against calibration error. Furthermore, migration experiments on a self-built multimodal dataset confirm its strong generalization capability, establishing it as a practical and efficient new method for industrial deployment.
Ying H. Ma, Che Liu, Jia-Jing Wu et al.· IEEE Transactions on Radar S...· 0 citations
Millimeter-wave (mmWave) radar has found extensive applications owing to its superior penetration capability and high-resolution imaging performance. In conventional near-field mmWave imaging, meeting the spatial sampling requirements for high-resolution imaging typically necessitates the deployment of oversized antenna arrays or densely spaced transceiver sampling, which leads to increased data acquisition time, significantly elevated data processing complexity, and higher system costs. To overcome this limitation, we present an imaging approach based on a flow matching (FlowM) generative model, termed FlowM-mmWImager, which enables high-quality target reconstruction by directly processing the 16-times undersampled raw radar echo data acquired from a small-aperture antenna array. By incorporating physical constraints derived from target electromagnetic scattering information, FlowM-mmWImager employs a neural-network-modeled velocity field to smoothly transform random Gaussian noise into the target image distribution, achieving high-fidelity radar image generation. To validate the effectiveness of our method, we construct two simulation datasets, SimSAR-MNIST and SimSAR-HWDB, for training and evaluation, and further develop an additional test set, SimSAR-EMNIST, to assess generalization capability. For experimental validation, real-world measurement data comprising common metallic tools are collected using our self-developed mmWave radar system to evaluate the method’s performance in practical scenarios. Experimental results demonstrate that FlowM-mmWImager yields imaging results highly consistent with ground truths in both simulated and real conditions, with an imaging time of less than 1 s, exhibiting strong cross-dataset transferability and practical value.
Huan Zhang, Che Liu, W. Yu· IEEE Transactions on Instrum...· 0 citations
We use cookies to run the site and, with your consent, for analytics and to show ads.
See our Cookie Policy.
