Research On Multi-source Information Fusion Indoor Positioning Method Based On Scene Recognition With NetVLAD

With the acceleration of urbanization and the development of the Internet of Things,the production mode is shifting more towards indoor environments and people’s daily activities are increasingly taking place indoors,leading to a sharp increase in the demand for indoor positioning services.Perceiving the user’s indoor environment is of great significance for providing intelligent and reliable location services to the user.Compared to outdoor scenes,indoor scenes have obvious spatial restrictions,relatively fixed structural layouts,and contain a lot of invalid information such as walls and floors,posing great challenges to scene recognition.Due to the closed environment of indoor spaces,GNSS technology cannot provide effective indoor positioning services,which has led to the emergence and development of various indoor positioning technologies.There are various interferences in indoor environments,such as signal attenuation,multipath effects,and non-line-of-sight transmission,which affect the accuracy and reliability of indoor positioning technologies.This makes it difficult for indoor positioning methods based on a single positioning source to meet positioning needs.However,different positioning technologies have their own strengths and weaknesses,so fusing information from multiple positioning sources is an effective way to improve positioning performance.Focusing on the needs of scene recognition and high-precision positioning in indoor environments,this thesis proposes a multi-source information fusion indoor positioning method based on scene recognition,integrating three positioning technologies: image vision,Wi Fi,and PDR.The main work includes:(1)We studied the scene recognition process based on Net VLAD.Given the importance of image features in scene recognition,we compared the accuracy of scene recognition using different feature extraction networks and determined Res Net as the backbone network for image feature extraction in Net VLAD.Considering the fact that indoor scenes usually contain a large amount of useless information such as walls and floors,and the scene images are captured at different distances,angles,and lighting conditions,we added the ASPP＿CBAM module to expand the receptive field and extract multi-scale features,as well as to make the network more attentive to important scene features.This enhanced the network’s feature extraction ability and improved the accuracy of scene recognition.(2)We proposed a visual image-based indoor positioning method based on multimatching range constraints.We used SIFT features for image matching,and optimize the matching results through multiple constraints.To solve the problem of low efficiency in image feature matching over a large range,we constructed a candidate matching image feature library through twice screening of Wi Fi fingerprint matching results and heading information,narrowing the matching range and achieving fast and accurate matching.The terminal position cannot be estimated directly due to lacking scale information in pose estimation using epipolar geometry.To address this issue,we converted the pose results obtained by multi-matching images into relative directions between the current image and the database images,and then the terminal position is estimated using least squares method with spatial intersection to achieve high-precision image visual positioning.(3)We studied pedestrian gait detection and step length estimation models based on smartphone inertial sensor data,and combined them with heading angle calculation to implement PDR positioning method.To address the problem of time drift in PDR,we used an Unscented Kalman Filter to fuse the relative PDR positioning results with absolute positioning results from visual and Wi-Fi sources.This corrected the PDR results and also reduced the jumping phenomenon in the visual and Wi-Fi positioning results.During the filtering fusion process,an adaptive measurement update strategy is chosen based on the position relationship between the predicted state and the visual and Wi-Fi positioning results.Additionally,the measurement noise is adaptively estimated to improve the positioning accuracy and robustness of the system,providing navigation positioning results that are close to the real trajectory.The thesis contains 59 figures,10 tables and 86 references.