Research on indoor positioning in underground parking based on multimodal sensors and deep learning: model optimization and performance analysis

This study explores indoor positioning systems by integrating multimodal sensor data and employing advanced deep learning architectures, including MLP, LSTM, BiLSTM, and CNN+MLP models. It addresses challenges posed by complex indoor environments, such as signal noise, multipath interference, and dynamic movement patterns, which frequently degrade localization accuracy. A comprehensive performance evaluation was conducted to assess the models' accuracy, robustness, and computational efficiency. Experimental findings reveal that the CNN+MLP architecture outperformed other models, achieving an average mean absolute error (MAE) of 5 meters. It demonstrated enhanced stability and generalization capabilities, especially in scenarios with increased noise and dynamic movements. In contrast, LSTM and BiLSTM models exhibited higher errors due to difficulties in adapting to temporal dependencies under variable conditions. The CNN+MLP model effectively captured spatial features, providing superior performance for static and low-motion environments. This research highlights the significance of integrating spatial and temporal features to improve positioning accuracy in complex indoor environments. Key deployment challenges, such as computational overhead, hardware limitations, and scalability, are discussed in detail. The study also proposes future enhancements, including the incorporation of attention mechanisms, lightweight model optimization for mobile applications, and transfer learning strategies to improve generalization across diverse environments. These findings lay the groundwork for advancing high-precision, real-time indoor positioning technologies.