Research On Key Technologies Of Inertial Navigation-based Pedestrian Dead Reckoning

With the acceleration of urbanization,people have an increasing demand for location services.A global navigation satellite system(GNSS)can almost meet people’s needs in outdoor areas.However,due to the attenuation and occlusion of satellite signals,GNSS cannot achieve full coverage in an urban canyon,indoor environment,and underground spaces(such as subway,underground commercial street,underground corridor,etc.).However,these environments cover a range of public activities from 70% to90%.Therefore,to ensure the safe travel of visually impaired people,there is an urgent need to study flexible navigation and positioning technology suitable for all city spaces,with substantial autonomy,anti-interference ability,and high precision.With the miniaturization,performance improvement,and cost reduction of selfcontained sensors(including accelerometer,gyroscope,and magnetometer),pedestrian dead reckoning(PDR)technology based on inertial navigation becomes a solution for full space positioning(indoor and outdoor seamless positioning)with substantial autonomy,high reliability,and excellent performance.It is gradually attracting the attention of academia and industry.Regarding the two core problems of heading drift and position drift of the PDR system,the main innovations and research results are summarized as follows:(1)long-term heading drift correctionThe basic PDR framework based on an inertial navigation system(INS)only utilizes angular rate for continuous attitude calculation,which is easy to produce a cumulative error.Meanwhile,the heading error cannot be corrected by the zero-velocity update(ZUPT)algorithm.Therefore,with the increase in operation time of the PDR system,heading drift is easy to occur,which eventually leads to positioning deviation.However,the robustness of current algorithms becomes weak,and their performance degrades with complex environments and human motions.Therefore,an adaptive heading fusion correction method based on magnetic disturbance detection and human motion recognition is proposed and integrated into the basic PDR framework(ZUPT-AFM for brief).On the high-precision visual capture Vicon public dataset with a total distance of 324.3 m,the heading drift of the ZUPT-AFM framework is 0.83°/minute,which is56.7% lower than that of the basic PDR system.On the indoor and outdoor motion dataset with a cumulative distance of 12.98 km,a comparative experiment is carried out with PDR frameworks based on mainstream heading correction algorithms.The ZUPT-AFM framework achieves mainstream accuracy.Namely,the 75% total traveling distance error(TTDE)is less than 0.59%.100% TTDE is less than 1.22% is 60.8%lower than the maximum TTDE of the basic PDR framework.(2)zero-velocity detection under the dynamic gait speedThe position drift of the PDR framework based on INS will increase rapidly without any constraints.Because the zero-velocity phase exists in a gait cycle,the ZUPT algorithm is generally used to correct the position drift.Because the zero-velocity detector triggers the algorithm,the performance of zero-velocity detection directly affects the positioning performance of the PDR system.Current algorithms are robust enough,and their performance degrades for fast-changed gaits.Therefore,this thesis discusses the relationship between the threshold parameters of the classical zero-velocity detector(SHOE)and the gait speed.Then the zero-velocity detection problem is transformed into the threshold regression problem under the dynamic gait speed.In addition,an adaptive zero-velocity detector with robustness to the gait speed is constructed using statistical machine learning and deep learning technology combined with a SHOE detector.The experimental results show that the proposed deep learning model(TRNet)is better than the proposed statistical machine learning model,the traditional model based on the best threshold,and the the-state-of-the-art open-source adaptive zero-velocity detection model.Compared with the SHOE detector based on the optimal fixed threshold,the TRNet proposed in this thesis reduces the root mean square error(RMSE)of the basic PDR system distance error by 48.7%,the RMSE of the start-end error by 12.5%,and the average of RMSE of position errors by 19.2%.(3)long-term horizontal position drift correctionAffected by the cumulative error,the PDR system is prone to position drift,challenging to achieve accurate long-term positioning.To address this issue,hardware beacons,signal fingerprints,and maps are generally leveraged in current works.This thesis creatively uses the preset visual beacon points to correct the horizontal position error.Our proposed method does not require external hardware beacons and map information.Compared with fingerprinting methods,visual beacons have a longer maintenance cycle and lower difficulty to maintain.First,a lightweight visual beacon recognition neural network with strong robustness to light,seasons,and camera angles is proposed;a private dataset containing 50 different campus landmarks is established.The experimental results on private data sets and public Pitts30 k data sets show that the proposed model’s performance is close to the current mainstream visual beacon recognition neural network(VGG16-Net VLAD).Its floating-point operations(FLOPs)and parameters are only approximately 1% and 19.8% of VGG16-Net VLAD.Second,because captured images do not necessarily belong to the retrieval database,this thesis designs an open-loop visual beacon retrieval algorithm based on the proposed network model to exclude non-beacons.Finally,a Kalman filter method with distance constraints is proposed to correct the horizontal position drift error of the PDR system.Long-time and long-distance walking experiments show the effectiveness and high performance of the long-term horizontal position drift correction algorithm.Compared with the ZUPTAFM algorithm,the start-end error is reduced by 92.16% with our method,and 68.64%reduces the RMSE of position error.