Research On Key Technology Of Image And Wireless Signal Fusion Location Based On Deep Learning

With the popularity of mobile Internet,the upgrade of smart mobile devices and the arrival of 5G era,the demand for LBS(Location Based Services)is increasing.At present,the two most promising indoor positioning technologies are based on wireless signals and computer vision.These two kinds of technologies have their own advantages and disadvantages:wireless signal-based indoor positioning technology has the advantages of low computational complexity and strong realizability,but it has the disadvantages of multi-path,non-line-of-sight interference,and insufficient feature space discrimination;computer vision-based indoor positioning technology has the advantage of obtaining a large number of stable environmental information,but it has high computational complexity and is affected by illumination.Aiming at the limitation of single indoor positioning technology,this paper proposes a fusion positioning method of image and wireless signal based on depth learning.Firstly,the landmark detection model is used to detect the landmark in the image,and these landmarks are used to express the image information.The two-dimensional image information is mapped to one-dimensional features.At the same time,the wireless features are extracted by the weighted extraction model.Then,the fusion of landmark features and wireless features is realized by using the fusion location network.Finally,the location is estimated by regression method.This paper mainly completes the following four tasks:1.To solve the problem of how to fuse image information and wireless information in the same dimension,a landmark detection model based on convolutional neural network is proposed in this paper.The landmark detection model is used to detect the landmarks in the image,and these landmarks are used to express the image information,so that the two-dimensional image information can be mapped to one-dimensional landmark features.The feature is used to input the final fusion network model.2.Aiming at the problem of landmark selection,this paper proposes a landmark feature selection method.Because of the large number of targets,if all targets are regarded as landmarks,a lot of labeling work is needed,which makes the feasibility of the fusion positioning method reduced.According to the feature selection method proposed in this paper,the target features with strong independence are screened,and the target distribution is more uniform while the redundant targets are filtered.The average accuracy of target detection is improved from 0.69 to 0.90 in the case of small sample training.3.To solve the problem of how to obtain reasonable training data,this paper proposes a weighted extraction model.Due to the difference of WIFI information received by different mobile phones,in order to obtain robust training sets suitable for different mobile phones,this paper proposes a weighted extraction model to fuse WIFI data received by different types of mobile phones as training data.Experiments show that the average positioning error before and after fusion is reduced by about 0.3m when tested on another type of mobile phone.4.Aiming at the fusion of landmark features and WIFI features obtained by preprocessing,this paper proposes a fusion network model based on in-depth learning.The model first connects the two features separately,then cascades them,and finally calculates the location results by regression method.Compared with the better indoor positioning methods,the average positioning error of the fusion model is reduced by about 1 m in the experimental scene of this paper.Finally,through experimental analysis and comparison,it is proved that the method proposed in this paper has the advantages of realizability and positioning robustness for different types of mobile phones in multiple scenarios.