ACCOR: Attention-Enhanced Complex-Valued Contrastive Learning for Occluded Object Classification Using mmWave Radar IQ Signals

Millimeter-wave (mmWave) radar provides robust sensing under adverse conditions and can penetrate thin materials for non-visual perception in industrial and robotic settings. Recent work with MIMO mmWave radar has demonstrated its ability to penetrate cardboard packaging for occluded object classification. However, existing models leave room for extensions and improvements across different sensing frequencies. Building on recent work with MIMO radar for occluded object classification, we propose ACCOR, an attention-enhanced complex-valued contrastive learning approach for radar, enabling robust occluded object classification. ACCOR processes complex-valued IQ radar signals via a complex-valued CNN backbone, a multi-head attention layer and a hybrid loss. The hybrid loss combines a weighted cross-entropy term with a supervised contrastive term. We extend an existing 64 GHz dataset with a new 67 GHz subset and evaluate performance across both bands. ACCOR achieves 96.60 % accuracy at 64 GHz and 93.59 % at 67 GHz on 10 objects, surpassing prior radar-specific and adapted image models. Results demonstrate the benefits of integrating complex-valued deep learning, attention, and contrastive learning for mmWave radar-based occluded object classification.

💡 Research Summary

The paper introduces ACCOR, an attention‑enhanced complex‑valued contrastive learning framework for classifying objects that are hidden inside cardboard packaging using millimeter‑wave (mmWave) radar IQ signals. The authors build on prior work that demonstrated the feasibility of using a compact 20 × 20 MIMO FMCW radar (62‑69 GHz) to penetrate thin materials and recognize the contents of a sealed box. While earlier approaches either relied on computationally heavy 3‑D convolutions or transformed the complex IQ data into real‑valued images (range‑Doppler, range‑angle), ACCOR processes the raw complex‑valued data directly, preserving both amplitude and phase information.

The architecture consists of three main components. First, a complex‑valued convolutional neural network (CNN) backbone receives the FFT‑transformed IQ matrix (400 virtual channels × 100 range bins). Complex convolutions, batch normalization, and a complex ReLU (cReLU) are employed, ensuring rotational invariance and magnitude‑phase coupling are retained. After three convolutional layers with 5 × 5 kernels and average pooling, the network outputs a 256‑dimensional feature vector. Second, this vector is split into real and imaginary parts, concatenated, and fed into a multi‑head self‑attention (MHSA) layer with 16 heads. The attention mechanism captures long‑range dependencies across range and angular dimensions, refining the representation before classification. Third, training uses a hybrid loss: a weighted cross‑entropy term (ℓ_χ) combined with a supervised contrastive term (ℓ_κ) weighted by a factor α. The contrastive loss treats each sample as an anchor, pulls together features of the same class (positives) and pushes apart features of different classes (negatives) using a temperature τ, thereby increasing inter‑class separability—a crucial advantage given the high similarity of radar signatures.

To evaluate frequency‑dependent performance, the authors extend the original 64 GHz dataset (from