Key Research On Indoor Localization Algorithm Using Image And Machine Learning Technique

With the increasing requirement of location based service,image based indoor localization has become a research hotspot.With the development of machine learning,indoor image based localization can be formulated as a theoretical machine learning model.In this paper,the key technologies of indoor localization using image and machine learning are studied.The main contributions are described as follows:First,an image based localization algorithm using image similarity measurement and back propagation neural network(BPNN)regression learning is proposed.In the offline stage,three image similarities(structural similarity,histogram similarity and cosine similarity)and relative distance between the collection position and the reference position are calculated as the fingerprint and label of the training data.Then the BPNN is used for regression learning and the distance based regression function is obtained.In the online stage,after the similarity between the received image and the reference image is calculated,the distance is estimated by distance regression function.Experiment results show that the proposed algorithm has better localization performance than existing algorithms.Second,a multi-image localization method based on Pix2 Pix image transformation strategy is proposed for the absence of collected image.In the offline stage,the missing image is supplemented by Pix2 Pix network.Then,multiple images are spliced to form training data.At last,the convolutional neural network(CNN)is used for image classification learning and the position based classification function is obtained.In the online stage,the position based classification function is used to realize location.Since the Pix2 Pix network is based on the conditional generates adversarial network and uses the combination of L1 distance and adversarial loss,the generated image can be closed to the source image.Experiment results show that the proposed algorithm can effectively solve the problem of missing image,so as to obtain better positioning performance.Third,a multi-image localization method based on Embrace Net multi-mode fusion is proposed for the absence of collected image.In the offline stage,CNN is firstly used to extract the features of the acquired images,then Embrace Net is used for image feature fusion,and finally the Soft Max layer of CNN is used for classification learning and get the position based classification function.In the online stage,position based classification function is used for position estimation.In Embrace Net fusion model,the information loss of the missing image modality can be overwritten by other image modalities which can ensure the robust localization performance for missing image.Experiment results show that the proposed algorithm can achieve good localization performance even when the image is missing.