A Data-Driven Method for Indoor Radar Ghost Recognition With Environmental Mapping

Abstract

Millimeter-wave (mmWave) radar has been widely applied in target detection. However, due to multipath and occlusion, radar often detects ghosts, especially in indoor environments. Existing solutions are mostly tailored to specific, simplified scenarios. To identify radar ghosts in diverse and complex indoor environments, we propose a data-driven approach. A thoughtful indoor radar ghost dataset is created with a multimodal data acquisition and automatic annotation system. And PairwiseNet, an end-to-end deep neural network adept at handling point-pair relationships within sparse point clouds, is proposed for radar ghost recognition. Multiframe accumulation is also implemented in PairwiseNet. To further enhance PairwiseNet, an additional network incorporating grid maps and U-Net is developed for constructing environmental maps from sequential point clouds. This network is trained through cross-modal distillation, with a depth camera as the teacher. Finally, a series of experiments validates the effectiveness of the proposed method in identifying indoor radar ghosts and autonomously constructing environmental maps. The classification accuracy on the test set reaches 96.0%, accurately identifying ghosts in the vast majority of cases.