A Multi-scale Spatial-temporal Features Fusion Framework for Indoor Localization

WiFi positioning technology has attracted considerable attention in recent decades due to its widespread deployment and cost-effectiveness. The multipath effect can lead to different local variations in WiFi signals, diminishing both localization accuracy and robustness. In this paper, we present an innovative localization framework that employs multi-scale spatial and temporal features for localization, which takes the Received Signal Strength (RSS) sequence as input. First, we propose a multi-scale spatial feature extraction network to capture multiple local features by using different convolutional operations. Then, a deep temporal network based on the Gate Recurrent Unit (GRU) is used to explore signal correlations at the temporal level. Finally, a channel-spatial (CS) attention mechanism is applied to discriminate the importance of multi-scale spatial and temporal representations. Guided by the acquired attention values, multiple features are fused to generate more discriminative representations for localization. Extensive experiments are conducted to validate the effectiveness of our scheme, and the results demonstrate its superior localization accuracy and robustness compared to other localization approaches.