Research On Data Fusion Algorithm For Pedestrian Dead Reckoning

The emerging high-accuracy indoor positioning technology for pedestrian has become an important research direction.At present,pedestrian dead reckoning technology can be divided into two categories:active positioning and passive positioning.Among them,the passive positioning system does not depend on the external environment and is an autonomous positioning system.Its advantages lie in low cost,high concealment,strong anti-interference,and potential for full scene localization,especially in solving high precision positioning problems in scenarios without satellite signals.However,since this type of system estimates the relative pose transformation of the target,this leads to accumulation of errors causing challenges in performing accurate indoor positioning.Therefore,leveraging the wearable devices,this thesis focuses on the study of a multi-integrated pedestrian dead reckoning system that integrates inertial data,pedestrian movement patterns and image data for real-time positioning calculations in unknown scenarios.The core researches contain:The strapdown inertial navigation system assisted by the ZUPT is built that integrates pedestrian movement patterns and inertial data,that is,a zero velocity update assisted pedestrian dead reckoning system.To fulfill the human gait classification,the deep learning model(i.e.,Symmetrical-Net)is used for optimization,which improves the adaptive detection compared with the conventional general likelihood ratio test.Dead reckoning system performance is also improved by the addition of the adaptive detector.In addition,this thesis also collected a set of data sets in real scenarios for the training the model.The visual-inertial navigation system is built that integrates inertial data and image data.A tightly coupled fusion model based on a neural network(i.e.,DenseVIO)is designed.By extracting a high-dimensional representation of sensor information,learning the contextual information,the end-to-end estimation model trained on the open-source dataset is proposed in this thesis.The problems of robustness and low precision of positioning systems that depend on a single source are alleviated.