HQCB-DETRLoc: a DETR indoor localization method based on HHO-QGA hybrid optimization and CBAM attention

Abstract Device-free localization (DFL) is an emerging internet of things (IoT) sensing technology for locating targets without attached devices. The core challenge lies in extracting effective features from weak signal fluctuations to achieve high accuracy and robustness. To address this, the paper proposes HQCB-DETRLoc, a novel indoor localization method that integrates a Modified Residual Network with a Convolutional Block Attention Module (ModifiedResNet-CBAM), a detection transformer (DETR), and a hybrid Harris Hawks Optimization-Quantum Genetic Algorithm (HHO-QGA). The method employs the channel and spatial attention modules of ModifiedResNet-CBAM to concentrate on informative positioning features within channel state information (CSI) signals. A heterogeneous fusion structure accommodates the dimensionality and semantic heterogeneity of multi-scale features, thereby enhancing feature extraction. The HHO-QGA deep fusion optimizer intelligently searches for optimal hyperparameters by synergizing the hunting strategy of Harris hawks with quantum state representation. A reparameterization block (ReparamBlock) is utilized to balance training complexity and inference efficiency. Meanwhile, a TransformerEncoder captures cross-space global dependencies of CSI features. A two-stage training framework decouples hyperparameter search from model parameter optimization.By integrating attention enhancement, heterogeneous fusion, and global modeling, HQCB-DETRLoc exhibits good robustness against extreme noise and anomalies in DFL scenarios. Experimental results on real-world datasets show that HQCB-DETRLoc achieves an average localization error of only 0.44 m. Compared to state-of-the-art methods, HQCB-DETRLoc not only delivers superior localization accuracy but also accommodates real-time deployment requirements via its reparameterization module, establishing it as an efficient and high-precision indoor positioning solution.