HG-Net: a novel neural network with hierarchical grouped convolution for indoor fingerprint positioning

Deep learning-based methods have shown promising performance in CSI-based fingerprint localization. However, current approaches often adopt feature interaction across different base stations (i.e., channels) throughout the entire process to achieve base station modeling. This approach conflates inte...

Full description

Saved in:
Bibliographic Details
Published in:Cluster computing 2025-04, Vol.28 (2), p.122, Article 122
Main Authors: Meng, Xiangxu, Li, Wei, Liu, Zhihan, Yang, Junze, Zheng, Wenqi
Format: Article
Language:English
Subjects:
Accuracy
Algorithms
Computer Communication Networks
Computer Science
Deep learning
Feature extraction
Fingerprints
Global positioning systems
GPS
Localization
Machine learning
Methods
Neural networks
Operating Systems
Processor Architectures
Recurrent neural networks
Semantics
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Deep learning-based methods have shown promising performance in CSI-based fingerprint localization. However, current approaches often adopt feature interaction across different base stations (i.e., channels) throughout the entire process to achieve base station modeling. This approach conflates interactions between antennas within the same base station and those between different base stations, leading to inefficiencies. Additionally, processing each CSI point as a separate unit introduces a significant computational burden. In this work, we propose a new neural network, called HG-Net, which can use single-channel features to obtain a more generalized learning representation and achieve better localization performance. Specifically, HG-Net adopts a Patchify for coarse-grained feature extraction. Then, Single Channel Feature Extractor extracts feature from a single channel, while Hierarchical Feature Aggregation stores three semantic features of different granularity and merges them to obtain rich deep features. In addition, a Recurrent Neural Network is developed to obtain long-range frequency domain dependencies. Finally, the linear layer-based Positioner can map the deep features to locations. Experimental results on three real-world datasets show that our proposed HG-Net achieves an average positioning deviation of 0.14 m, which greatly exceeds the state-of-the-art approaches.
ISSN:1386-7857
1573-7543
DOI:10.1007/s10586-024-04793-w