An indoor positioning algorithm fusing stacked auto-encoders and residual networks

Abstract

To enhance localization precision in multi-story complex environments, we propose an innovative indoor localization methodology that integrates an advanced stacked auto-encoder with a refined residual neural network. Initially, we architect a convolutional branch featuring multi-layer convolutional operations, built upon the foundation of a stacked auto-encoder, to facilitate skip feature connections. Subsequently, we introduce an optimized residual neural network that employs a one-dimensional convolutional layer for feature extraction, utilizes global average pooling for dimensionality reduction, and amalgamates the output of global average pooling with a multi-layer fully connected layer. This design ensures the extraction of deep features while preserving critical feature information, thereby mitigating the gradient vanishing issue prevalent in deep networks. Furthermore, our study incorporates an enhanced genetic algorithm to explore the global optimal solution, thereby augmenting the accuracy of indoor positioning. Empirical results on the UJIndoorLoc public dataset demonstrate that our proposed method achieves an average positioning error of 7.85 meters, with building identification accuracy reaching 100% and floor positioning accuracy attaining 97.2%. These results signify a substantial improvement over existing algorithms, rendering our approach particularly suitable for scenarios demanding higher positioning accuracy.