NLOS Signal Recognition And PPP-RTK Optimization Based On Vision And Deep Learning

PPP is one of the main technologies for GNSS to provide high-precision positioning services.It has the advantages of flexible single operation and high positioning accuracy.However,traditional PPP technology still has the problem of long convergence time,which can not meet the needs of dynamic applications.In order to overcome the limitations of PPP,some scholars put forward PPP-RTK technology.By extracting accurate atmospheric corrections from the ground reference network,regional users can quickly fix the ambiguity and obtain centimeter-level positioning accuracy in a few seconds.However,despite the rapid development of GNSS positioning algorithm,its positioning performance is still seriously limited by the environmental conditions,especially in the complex urban environment,non-line-of-sight(NLOS)signals greatly affect the performance Of GNSS.Therefore,in order to achieve accurate and rapid positioning in the urban complex environment and suppress the influence of NLOS,this thesis focuses on the NLOS signal recognition and PPP-RTK positioning optimization method in the complex urban environment.The main work and conclusions of the thesis are as follows:(1)Based on dynamic vehicle-mounted data in typical city scenes,the performance of PPP-RTK under different observation conditions are evaluated and analyzed,and the influence of NLOS signal on PPP-RTK in GNSS semi-occluded environment is discussed.The results show that the positioning accuracy of PPP-RTK is stable at cm level and the fixed rate is higher than 95% in the open field of vision,but the positioning performance of PPP-RTK is significantly reduced in the semi-occlusion scene,the positioning accuracy is reduced from cm level to decimeter level,and the fixed rate of ambiguity is decreased.The fixed rate decreased to 88.95%,74.49%,71.69% under urban canyon,shady forest and bridge occlusion.The location stability and reliability decreased,and the flying points and gross errors increased.Furthermore,this thesis analyzes the influence Of NLOS signals by comparing the observation data and positioning results in line-of-sight(LOS)and NLOS environments.The results show that NLOS signal can reduce the receiving intensity and signal quality of the receiver,and increase the incidence of pseudo-range noise and cycle slips.In terms of positioning results,NLOS signal introduces a large range error,resulting in poor precision of floating point ambiguity,which makes the ambiguity easy to fix failure or fix error,and the positioning accuracy deteriorates or the positioning sequence appears coarse error.(2)Taking advantage of the visual sensor’s ability to effectively perceive the geometric distribution of surrounding occlusion,a method of NLOS signal recognition is proposed based on sky images taken by fish-eye camera.GNSS observation equipment,inertial sensor and fisheye camera are combined to obtain the environmental occlusion and the communication relationship between the satellite and the receiver.Firstly,the scene model is constructed,and an effective method to segment the occlusion area and sky area in fisheye image is proposed.Then,the image is corrected based on the pose information provided by IMU sensor,and the satellite coordinate projection model is constructed.Finally,the satellite visual situation is analyzed and judged based on the scene model.(3)The PPP-RTK method is improved and optimized based on NLOS identification information,and the influence of NLOS ia suppressed in terms of the observation and the fixed ambiguity to improve the anti-multipath performance of PPP-RTK.In typical city scenarios,the positioning performance improvement effect of the optimized PPP float solution and PPP-RTK fixed solution are evaluated.The experimental results show that for the geodetic receiver,91% scenes’ positioning accuracy is improved after PPP float solution optimization.In the case of the urban canyon and bridge,the positioning error decreases from 7.08 m and 5.90 m to 3.71 m and 2.89 m,respectively,with the accuracy improvement of 47.9% and 51.0%.For the low cost receiver,82%scenes’ positioning accuracy is improved,and the improvement effect is most obvious for the continuous bridge and shady scenes with serious occlusion.After the PPP float solution experiment proves the effectiveness of weakening the influence of NLOS signal,the optimization effect of positioning performance of multi-constellation PPP-RTK fixed solution is further explored combined with the fuzzy fixed optimization strategy.The experimental results show that the positioning accuracy of 64% scenes is improved after optimization for the geodetic receiver,and the positioning errors of bilateral occlusion,urban canyon occlusion and bridge occlusion with obvious improvement are reduced from 1.69 m,1.89 m and 1.41 m to 1.42 m,1.60 m and 1.02 m,respectively.The corresponding accuracy increases are 16.0%,15.3% and 27.7%,respectively.For the low cost receiver,73% scenes’ performance is improved,and the performance of continuous bridge,urban canyon and shady forest scene is significantly improved.(4)Although the NLOS recognition method based on fish-eye image is effective,its practicability is limited due to the need for external equipment.Therefore,combining NLOS information with GNSS eigenvalues is considered to study the deep learning based NLOS signal recognition method.In the thesis,an end-to-end LOS/NLOS signal classification network with nonlinear modeling capability is designed,and the classification model is trained and evaluated based on urban vehicle-mounted data.Meanwhile,the classification results of support vector machine(SVM)and random forest(RF),two traditional machine learning methods,are compared.The results show that the classification performance of neural network is better than that of machine learning,and the classification accuracy,accuracy and recall rate are 88.3%,80.6% and 59.2% respectively.Furthermore,the NLOS signals identified by neural networks are applied to PPP-RTK positioning solution to verify the correctness and availability of the identification results.The PPP-RTK solution experiments are carried out for the four test data sets respectively.The results show that the corresponding solution results of NLOS signals identified by vision and neural network are roughly the same,and the positioning accuracy and stability are better than the PPP-RTK results without considering the NLOS error.In the case of urban canyon,bilateral occlusion and bridge occlusion,the corresponding positioning accuracy of neural network NLOS recognition results is improved by 15.8%,11.1% and 42.8% respectively,which proves the effectiveness and availability of the signal classification model.