Research And Design Of Multi-source Fusion Indoor Positioning System Based On Scene Recognition

With the development of Io T and the progress of society,there is a growing demand for indoor environmental services.Understanding the user’s inside environment makes it easier to provide them with tailored and intelligent location services.However,indoor environment is frequently complex and changing,posing significant hurdles in scene recognition and navigation.There are relatively few studies combining indoor scene recognition and indoor positioning,and most indoor positioning methods rely on additional devices.It is of great scientific significance and commercial value to propose a convenient and universal indoor positioning method applicable to indoor scene understanding.Based on the research of indoor scene recognition,this dissertation further investigates the smartphone localization methods based on vision,Wi-Fi and inertial sensors by combining the user’s demand for location services,and the main work and results of the dissertation are as follows:(1)Using transfer learning,three deep convolutional neural networks based on Image Net pre-training are used as feature extractors,and the extracted image features are selected with appropriate dimensions for dimensionality reduction,and fed into the classifier for classification training.Finally,integrated learning is used to combine the results of the models to obtain 90.6% accuracy of indoor scene classification.(2)Based on the results of indoor scene recognition,the indoor scene is divided into regions using block segmentation,and visual localization based on multi-view images of smartphones is implemented in the dissertation through a graphical localization network to reduce the user’s location information from the scene to the region and then further to the location coordinates.(3)Based on the visual localization,Wi-Fi localization algorithm,PDR(Pedestrian Dead Reckoning)localization algorithm and fusion localization algorithm are further analyzed.The dissertation finally proposes to use visual localization as the initial point and fuse Wi-Fi and PDR localization methods using an extended Kalman filter algorithm to achieve location tracking while the user is on the move.Experiments show that the fusion localization scheme has integrity and can effectively improve the problems of single localization method,and improve the localization accuracy by 14.5% and reduce the average localization error to 0.729 m compared with the common extended Kalman filter fusion localization method.In summary,the dissertation combines indoor scene recognition with indoor localization to achieve a complete,smartphone-side localization system without additional equipment,from scene recognition to area determination to initial point localization to final location tracking.