Deep learning for weights training and indoor positioning using multi-sensor fingerprint

• Main Authors: Xingli Gan, Baoguo Yu, Lu Huang, Yaning Li
• Format: Conference Proceeding
• Language: English
• Subjects: Data models; deep learning; Fingerprint recognition; indoor positioning; Machine learning; modeling data; multi-sensor fingerprint; Ray tracing; Training; Weight measurement; weights training; Wireless fidelity
• ISSN: 2471-917X
• DOI: 10.1109/IPIN.2017.8115923

Due to the influence of indoor signal multipath effect and human disturbance, the indoor positioning technology of WiFi fingerprint based on deep learning is poor stability. The large sample and accurate data in the room is very difficult to collect for weights training of deep learning, so it is difficult to be widely used. Firstly, the innovative algorithm with multi-sensor fingerprint and deep learning for indoor position (DL-IMPS) is puts forward, and used the statistical model and the ray tracing method to construct a large sample data for weights training, the experiment is proved that the model data of WiFi-RSSI is a subset of the actual measurement data. Secondly, 10.9m×7.4m indoor location test environment is set up in the room, through 9700 groups of modeling data and 1300 groups of measurement data to train DBN's weights, it get more optimal weights matrix and speed up the convergence rate. Finally, The performance of WKNN and DL-IMPS is compared under four different paths, The results prove that the average error of DL-IMPS is 0.52 m, the probability of error less than 1 m is 92.3%, but the average error of WKNN is 1.39 m, the probability of error of less than 1m is 45%, Location accuracy and stability of DL-IMPS are superior to WKNN. The other experiment is the comparison between one-sensor indoor location and DL-IMPS, Locating error probability of DL-IMPS is 1%, and the convergence speed is fast, that of WiFi-only is 24%, iBeacon-only is 25%, Geomagnetic-only is 15%, DL-IMPS have better positioning accuracy and robustness.