Geomagnetic indoor localisation-based dilated convolution neural networks

With the rapid development of mobile Internet, the demand for Indoor Location-Based Service (ILBS) keeps rising. Geomagnetic field fingerprinting-based positioning scheme, with low cost, high accuracy and good stability, has gradually earned the attention of researchers. There are many problems in the existing indoor localisation researches based on geomagnetic fingerprint recognition, such as heavy workload of fingerprint collection, insufficient positioning accuracy, high walking cost of positioning and long positioning delay. Therefore, a localisation system named DCGIL is proposed in this thesis. In DCGIL, a segmentation length estimation algorithm of geomagnetic fingerprint sequence is proposed to evaluate the minimum segmentation length, the shortest walking distance. Then, through an overlapping fingerprint segmentation method, the gap between coordinate points and the cost of fingerprint matching are reduced. Further, dilated convolutional neural networks is utilised for segmented fingerprint classification, which could improve the accuracy of fingerprint classification compared with the existing classification methods. Finally, DCGIL utilises the sliding window to align and match the segmented fingerprint. Experiments show that the localisation accuracy reached 1.21 metres at 3.8 m shortest walking distance derived from segmentation length estimation algorithm, which improved 28% compared to existing algorithms.