Boosting FMCW Radar Heat Map Object Detection With Raw ADC Data

Millimeter-wave (mmWave) radar is crucial for environmental perception in autonomous driving, especially under complex conditions. While radar heatmaps provide richer information than point clouds, extracting semantic details from heatmaps alone remains challenging. To address this, we propose leveraging raw radar Analog-to-Digital Converter (ADC) data and introduce Mamba-RODNet, a novel network that integrates radar heatmaps with ADC data. For long-sequence modeling such as ADC, Mamba outperforms Transformers in both accuracy and efficiency, making it well suited for autonomous driving perception. We further design an ADC-Mamba (AM) module that fuses multi-scale features from ADC and heatmaps, enhancing detection performance. Experiments on the large-scale RADDet dataset show that our method achieves state-of-the-art results in both average precision (AP) and floating point operations per second (FLOPs). Ablation studies demonstrate that incorporating ADC data improves mean Average Precision (mAP) by 7%. In summary, this work establishes a new paradigm for integrating raw mmWave radar ADC data into object detection, with significant implications for the field. Our code is available at here.