Research On Indoor Localization Method Based On Environment Depth Perception And Interaction

Guided by the development strategy of modernizing transportation information,the demand for indoor location services is growing.Wi-Fi and visual positioning signal sources have attracted much attention for their universality,low cost and easy deployment.Vision and Wi-Fi localization networks have problems such as measurement values susceptible to noise interference,poor performance of traditional localization algorithms,and defects of single-source localization networks.The thesis investigates these problems in terms of location feature depth extraction,high-precision localization algorithms,and multi-source fusion recognition and localization strategies.The details are described as follows.(1)Analyze the modulation and demodulation of Wi-Fi signals to model the CSI signal errors.The segmented phase calibration technique and the amplitude calibration and subcarrier screening techniques are used to reduce the influence of hardware devices on CSI amplitude and CIR.Meanwhile,the localizability of different fingerprints was analyzed.To design a reasonable location fusion strategy,one-dimensional CNN deep learning network is used to learn the mapping relationship between features and locations from the frequency and time domains.Then a decision fusion algorithm is used to estimate the target location under the same location space using high confidence prediction results and the prediction probabilities of the surrounding lattice points.(2)To reduce the influence of environment on image localization features.A lightweight YOLOv5 pre-trained network is selected to identify localization targets in the images and mitigate the effects of background changes.Based on the recognition results as well as the differences in pixel values and distribution differences for the same areas,the foreground features are extracted effectively and the interference of light and shadow is mitigated at the same time.The proposed support point-based target localization algorithm can accurately estimate the location in the image and surpasses most methods that only use target identification bounding boxes.In addition,based on the superiority of the framework,the positions of multiple classes of targets can be estimated.(3)To further improve the localizability of visual sources.To design a hardware framework for multi-source sensing information acquisition to achieve temporal and spatial synchronization of Wi-Fi and visual localization sources.Vision-based adaptive ?-?filters for tracking wireless localization trajectories.The discrete visual localization results are stitched together to generate anonymous visual trajectories based on multi-conditional constraints such as target motion state,time,space,heading,and speed.Finally,according to the Euclidean distance and similarity principle,MAC tags are bound for anonymous visual trajectories to realize the interactive recognition of multi-source localization networks.