PGA-TCN: Physics-Guided Attention and Temporal Convolutional Network-Based Millimeter-Wave Radar Ghost Removal

Abstract

This article addresses multipath-induced ghost targets in indoor millimeter-wave radar point clouds by proposing physics-guided attention and temporal convolutional network (PGA-TCN), which integrates motion-related physical priors with deep temporal modeling. A PGA module is constructed using adjacent-frame velocity variation and short-window spatial consistency descriptors, while a physical consistency loss is introduced as an auxiliary regularization term during training. Multiframe point-cloud sequences are processed through temporal convolutional networks to capture spatiotemporal dynamics. Experiments on the 1–3 pedestrian benchmark demonstrate that PGA-TCN achieves 95.12% accuracy with only 6.8 KB of parameters. Furthermore, additional evaluations are conducted in two previously unseen cluttered indoor environments with different spatial layouts and multipath propagation characteristics. The proposed method achieves 92.93% accuracy with a 93.52% macro F1-score in the four-person scenario and 92.16% accuracy with a 93.04% macro F1-score in the five-person scenario. These results demonstrate that PGA-TCN maintains effective ghost recognition capability and cross-scene robustness under challenging indoor radar sensing conditions, providing a lightweight solution for multipath ghost suppression and multiperson point-cloud classification.