Research on RFID Indoor Positioning Algorithm Based on Attention

Deep Learning‐Based Bluetooth Low‐Energy 5.1 Multianchor Indoor Positioning with Attentional Data Filtering

“…Tang et al [24] 0.2015 0.3657 0.2603 0.1973 0.3607 0.2592 MHSA-EC [22] 0 MSE and 37.3% MDE reduction in all cases compared to Kalman filtering). We also observe that as the number of time steps increases (more packets are used to form an input), the performances of average filtering and Kalman filtering on the two datasets also become better.…”

“…We compare the proposed AnFIPNet with the following methods: (MUSIC) algorithm; [ 33 ] PDDA; [ 18 ] MHSA‐EC; [ 22 ] Hi‐Loc; [ 23 ] attention‐based IPS model by Tang et al; [ 24 ] Gaussian–Bernoulli restricted Boltzmann machine plus liquid‐state machine (GBRBM + LSM); [ 34 ] time series attentional prototype network (TapNet); [ 20 ] CNN‐based joint APs model by Koutris et al; [ 9 ] and 2D image CNN by Hajiakhondi et al [ 11 ] For the MUSIC and PDDA algorithm, due to various interference and complex indoor environments, we remove 10% maximum and minimum phase difference outliers of each anchor to mitigate the phase difference fluctuations. After deriving the angle estimations from anchors, we use the least squares to estimate the final 2D positions.…”

“…Tang et al [24] 285.1081 0.0031 112.6071 0.0084 MHSA-EC [22] 452 AnFIPNet, with the designed feature, achieves the best performance in all metrics. More specifically, the performance of AnFIPNet with the designed feature is better than the performance with only phase differences, by an average of 41.5% MDE reduction in all cases.…”

“…Tang et al [24] 156.4610 0.0032 69.7960 0.0084 MHSA-EC [22] 153 [33] -0.1304 -0.1654 PDDA [18] -0.0573 -0.1131 GBRBM þ LSM [34] 3686.7621 þ 1737.7620 0.0247 3087.9884 þ 1102.1073 0.0235 Koutris [9] 56.5130 0.0011 40.6431 0.0009…”

“…Compared to them, our approach focuses on using the attention mechanism to filter out low‐quality data packets before feature embedding. Another study [ 24 ] proposed an IPS model that incorporates the self‐attention mechanism to filter the RSS, time of arrival, and angle information of the positioning tag. However, the self‐attention mechanism assigns weights to each feature based on its similarity with other input features.…”