Through-wall Radar Target Localization Method Based on Dual-stream Temporal Spatial Feature Extraction and DETR

Through-wall human target localization has broad application prospects in fields such as target perception and rescue. However, ultra-wideband through-wall radar systems suffer from wall clutter interference, which distorts target echo features and complicates the ability of traditional methods to achieve stable detection and high-precision localization in complex environments. Although deep learning-based localization methods have shown superior performance in these environments, they often rely on distributed radar layouts, leading to difficulties in system deployment and increased algorithm complexity. To address these challenges, this study introduces a deep learning network framework that utilizes a single-view small-aperture dual-transmitter quad-receiver ultra-wideband through-wall radar. This framework combines Dual-Stream Temporal Spatial (DSTS) feature extraction with a Detection Transformer (DETR) to accurately locate human targets behind walls. The network processes complex-range images as input, extracts spatiotemporal features, and constructs dual streams. The phase branch captures the target’s spatial angular information, and the amplitude branch reflects the target’s radial distance, thereby fully exploiting the distance and azimuth features in the echoes. The dual streams then undergo multi-scale downsampling, and a channel attention mechanism is employed for weighted fusion, yielding low-dimensional features. These features are then enhanced with positional encoding and fed into the DETR network, which utilizes its set-prediction capabilities to deliver reliable target localization results. Validation on measured data demonstrates that the proposed method achieves an average precision of 0.79, with a threshold for accurate multi-object localization set at 0.7 m, thus outperforming several existing solutions.