Temperature-induced hardware delay variations in Global Navigation Satellite System (GNSS) receivers degrade time synchronization accuracy in thermally unstable environments. This study introduces a novel methodology to characterize and compensate for temperature-dependent relative hardware delays between GNSS receivers, a critical limitation in high-precision timing. Employing a common-clock zero-baseline configuration with a programmable thermal chamber, we isolate receiver-specific delays by eliminating satellite-related errors, atmospheric effects, and external clock offsets. Experimental results across a 100 °C range demonstrate strong linear correlations (Pearson coefficients >0.97) between delays and temperature. Implementation of the compensation model reduced time synchronization errors significantly: standard deviation improved by > 61% (remaining < 0.38 ns), and RMS values fell below 0.63 ns. Stability analysis confirms effective suppression of thermal noise. The method, compatible with standard processing chains, demonstrates strong potential for enabling sub-nanosecond time synchronization in thermally variable environments, as evidenced by performance metrics that are statistically significant relative to the measurement uncertainty.

High-rate Global Navigation Satellite System (GNSS) is a valuable seismogeodetic tool for real-time monitoring of seismic displacements. The BeiDou PPP-B2b service provides precise point positioning (PPP) augmentation corrections via the B2b signal for BDS-3 and GPS satellites, while the Galileo High Accuracy Service (HAS) delivers such corrections for GPS and Galileo satellites via the E6B signal. However, PPP-B2b and HAS services are designed for different constellations, which makes the simultaneous use of BDS-3, Galileo and GPS impossible through a single service. To fully exploit the capabilities of multi-GNSS, we propose a method to integrate the PPP-B2b and HAS products with the emphasis on the positioning performance gain of the integrated products under high-rate and short-time seismogeodetic environment. In the simulated seismic wave experiments, PPP-B2b and HAS achieve positioning accuracies of 4.2 and 5.2 cm, respectively. For the 2024 Mw 7.0 Wushi earthquake, PPP-B2b attains accuracies of 0.50, 0.94 and 1.10 cm, while HAS achieves comparable accuracies of 0.83, 0.67 and 1.55 cm in the east, north and up components, respectively. Compared with standalone product, the integrated solution improves positioning accuracy by an average of 21 per cent in the single-axis shake table experiment and by 42 per cent, 29 per cent and 45 per cent during the Wushi earthquake, achieving average accuracy of 0.37, 0.58 and 0.69 cm in the east, north and up components, respectively. Magnitudes for the Wushi earthquake derived from the peak ground displacement of GNSS seismic waveforms using the PPP-B2b, HAS and integrated products show good agreement with the WUM product. These results confirm the feasibility of PPP-B2b, HAS and particularly their combined use for high-precision positioning, highlighting their great potential for real-time seismogeodetic applications in regions with limited communication infrastructure.

To fully exploit the structural characteristics of global navigation satellite system (GNSS) wideband multiplexed signals (WMSs) and enhance tracking robustness and positioning accuracy in complex environments, this article proposes a vector dual-assisted multicomponent tracking (VDMT) method. By integrating the subcarrier-level high-precision measurement of dual-assisted multicomponent tracking (DMT) with the global state feedback mechanism of vector tracking (VT), VDMT constructs a deeply fused tracking framework that unifies intrasatellite signal enhancement and intersatellite information cooperation. Architecturally, VDMT extracts and fuses asymmetric components from upper and lower sidebands to derive high-precision pseudoranges (PRs), fed into a navigation filter to jointly estimate the receiver's states across multiple satellites. The navigation solution is then fed back into each tracking channel to dynamically refine carrier, subcarrier, and code tracking parameters in a closed-loop manner. Experiments with BDS-3 B1 signals demonstrate that in static scenarios, VDMT reduces east/north position errors by 92%/83%, 89%/78%, and 44%/1% compared to conventional tracking (CT), VT, and DMT, respectively. In dynamic scenarios with signal attenuation, it maintains superior tracking stability and outperforms advanced DMT by reducing maximum and rms position errors by 82% and 11%, 33% and 3%, 89% and 41% in north, east, and up directions, respectively. Rather than merely combining the benefits of existing methods, VDMT realizes cooperative performance enhancement through structural integration, offering a new perspective for deep structural utilization and multichannel cooperative processing of WMS.

High-definition map (HD Map), as a critical component of digital infrastructure, plays a vital role in the advancement of autonomous driving technologies. Existing HD Map models are primarily structured according to the update frequency of map elements, and they insufficiently account for the service objects and task execution entities involved in autonomous driving. Moreover, these models often lack a systematic analysis of the interrelationships between map elements and their functional roles in autonomous driving scenarios. To address these limitations, this paper proposes a novel high-definition map model, termed the Human-Vehicle-Road-Map (HVRM) model, which considers the service objects, operational carriers, and supporting media of autonomous driving from the perspective of task execution. The proposed model integrates three core entities, namely humans, vehicles, and roads, into a unified high-definition map framework, and systematically analyzes the relationships among map elements and their significance in autonomous driving. The effectiveness of the proposed model is validated through simulation experiments. This research contributes to the existing body of HD map studies and offers new insights for the further development of autonomous driving systems.

Intelligent navigation is essential for unmanned systems. Yet nowadays navigation technologies still fall short of animals’ innate navigation prowess, characterized by continuous, efficient, adaptive, low-power navigating across complex terrains, and despite technological advancements. Neuroscience's half-century exploration has revealed the brain's innate “Global Positioning System (GPS),” instigating research into Brain-Inspired Navigation (BIN). BIN, is a cutting-edge navigation technology, that bridges disciplines but lacks a cohesive guide for its interdisciplinary study. In this article, we offer a comprehensive BIN review, mapping its neural basis, computational foundations, current progress, and implementation conditions, providing a general framework for researchers alongside forward-looking recommendations for future development in the domain. The highlights of this article can be available at https://binucoe.github.io/Awesome-Brain-inspired-Navigation/ .

Global navigation satellite system (GNSS) precise orbit, clock and observable-specific bias (OSB) products are a prerequisite for precise point positioning with ambiguity resolution (PPP-AR). Under the condition of sparse or regional ground network, integrating low earth orbit (LEO) onboard observations is a potential approach to supplementing GNSS global tracking measurements. In this contribution, a method of integrating ground and LEO onboard observations to generate GNSS precise satellite clock and phase OSB based on square root information filter estimation is proposed. In particular, a method of phase OSB estimation with a 1.5 h sampling rate is established using regional ground stations and eight LEO satellites. The results show that when integrating 30 global ground stations and eight LEO satellites, the standard deviation and convergence time of satellite clocks reach 0.093 ns and 4.93 h, which is a reduction of 41.5% and 35.4% compared to ground-only solutions, respectively. The phase OSB accuracy improves from 0.092 to 0.078 cycles, and the product availability increases from 42.2% to 95.1%. The kinematic PPP-AR accuracy of the integrated solutions is improved by 26%, 17% and 18% on the E, N and U components, and the convergence time is accelerated by 17%. In addition, the phase OSB products generated by 20 European regional ground stations integrating eight LEO satellites achieve an STD of 0.086 cycles, with a product availability of more than 98%, and the kinematic PPP-AR accuracy is close to that of solutions with phase OSB generated by 70 ground stations.

ObjectiveAccurate displacement prediction for reservoir-bank landslides is often hindered by two coupled difficulties. The first is the multi-factor forcing–response process driven by reservoir water-level fluctuations and seismic disturbances, where deformation responses are delayed and attenuated rather than instantaneous. The second is that many time-series approaches provide numerical forecasts but lack a physically interpretable indicator that can robustly capture turning points and support graded early warning under noisy, incomplete, or unevenly sampled monitoring data. To address these issues, a time-lag-aware displacement prediction and warning framework was developed for the Wangjiashan landslide in the Three Gorges Reservoir Area, with the aim of unifying the quantification of heterogeneous time-lag effects and strengthening the physical interpretability of the trend component used in forecasting and risk diagnosis.MethodsA displacement decomposition strategy was adopted to represent cumulative displacement as the sum of a trend component, a periodic component, and a residual term. The trend component was designed to reflect the long-term inertial creep and internal state evolution, and it was modeled using a low-order autoregressive structure. The periodic component was designed to capture the externally driven fluctuations associated with reservoir operations and seismic events, and it was formulated through time-lagged and memory-aware equivalent forcings. Time-lag quantification for reservoir water level was performed using grey absolute correlation analysis. Candidate lag days were scanned by shifting the water-level series relative to the displacement series, and the lag associated with the maximum correlation degree was adopted as the dominant response lag. To represent non-instantaneous and decaying hydrodynamic influence, an equivalent reservoir water-level series was constructed by applying a lag-dependent weighting scheme, so that recent water-level changes contributed more strongly than older ones within the effective lag window. Seismic influence was represented through an exponential convolution formulation that mapped discrete earthquake events into a continuous disturbance variable with short-term enhancement and decaying memory. Two parameters governed this mapping: an effective duration window and an exponential attenuation coefficient. These parameters were selected by maximizing the correlation between the constructed equivalent seismic variable and deformation response during the monitoring period. To connect predictive modeling with physically meaningful state diagnosis, a Trend Strength ratio was introduced. The ratio was constructed using sliding-window estimates of the velocities of the trend and periodic components after normalization. To reduce sensitivity to measurement noise, gap-filling jumps, and short-lived perturbations, Shannon entropy was used to quantify the dispersion of the velocity sequence within the sliding window, and an entropy-based suppression term was embedded in the ratio. An exponential moving average was then applied to stabilize the ratio update and prevent spurious oscillations. The sliding-window length was determined through a trade-off between noise sensitivity and temporal responsiveness, using a noise-amplification metric derived from multi-step differencing error propagation and an average-lag measure introduced by smoothing. Finally, the trend and periodic submodels were integrated into an Autoregressive Distributed Lag formulation, where displacement was predicted by combining autoregressive trend dynamics and time-lagged equivalent forcings.Results and Discussions For the Wangjiashan landslide, grey absolute correlation analysis identified an optimal reservoir water-level response lag of 8 days. After transforming raw water level into the equivalent water-level series using the identified lag structure, closer alignment was observed between the equivalent water-level variations and the displacement-rate evolution, supporting the necessity of time-lag incorporation rather than instantaneous forcing assumptions. For seismic disturbances, the exponential convolution formulation produced a best-performing parameter set with an effective duration of 5 days and an attenuation coefficient of 0.8 for the stronger event set. For weaker events, a shorter effective duration and faster attenuation were reflected by a smaller coefficient, consistent with the notion that stronger shaking leaves longer-lived mechanical and hydraulic signatures. The resulting equivalent seismic variable provided a continuous representation of post-event influence, avoiding the misinterpretation of isolated event points as trend shifts. For the Trend Strength ratio construction, the sliding-window length was set to 5 days based on quantified trade-offs. Increasing the window length from 3 to 5 substantially reduced noise sensitivity while introducing only a limited additional lag; further increases yielded diminishing marginal noise-reduction benefits while continuing to slow turning-point response. Shannon entropy played a clear physical role by distinguishing dispersion-dominated windows from coherence-dominated windows: windows dominated by noise, interpolation artifacts, or transient disturbances tended to show more dispersed velocity distributions and higher entropy, whereas sustained acceleration tended to produce more coherent velocity changes and lower entropy. The entropy-suppressed ratio therefore reduced false acceleration amplification and improved stability of state interpretation. The integrated time-lag-coupled ARDL model achieved improved prediction performance compared with a baseline model that ignored time-lag effects. Using the testing period specified in the manuscript, the proposed approach yielded RMSE of 26.59 mm, MAE of 20.68 mm, and R² of 0.998, whereas the no-lag comparison model produced RMSE of 37.44 mm, MAE of 33.00 mm, and R² of 0.9961. Error reductions of approximately 29% to 37% were obtained, indicating that explicit lag modeling contributed materially to predictive accuracy. Beyond numerical accuracy, the Trend Strength ratio supported state diagnosis and warning conversion. A graded warning scheme was defined using thresholds that separated stable, transitional, and accelerated stages. Stable behavior corresponded to Trend Strength ratio below 0.5, transitional behavior corresponded to values between 0.5 and 1.2, and accelerated behavior corresponded to values at or above 1.2. Under the monitoring period considered, three acceleration episodes were identified with a lead time of 1–2 days relative to the displacement-rate peaks, indicating that the ratio captured turning behavior earlier than direct rate-based indicators. Additional checks were conducted to examine robustness under practical monitoring constraints. When the monitoring frequency was reduced, predictive performance deteriorated markedly, highlighting that data discontinuity and sparse sampling can limit timely capture of rapid acceleration. In contrast, moderate random data loss was less damaging when the remaining time series preserved continuity, because displacement evolution retained temporal regularity that could be exploited by the model structure. Transfer tests using two independent monitoring points were also reported, where the predicted trajectories tracked observed cumulative displacement with high consistency in both amplitude and phase, and the associated performance indicators remained stable. These observations supported that the combination of time-lag quantification and ratio-guided trend updating can retain effectiveness beyond a single point series when the monitoring condition is comparable.ConclusionsA multi-factor time-lag coupling displacement prediction framework was developed by integrating displacement decomposition, grey-correlation-based reservoir lag identification, exponential-convolution-based seismic memory representation, and a Trend Strength ratio mechanism with entropy suppression and exponential smoothing. For the Wangjiashan case, the reservoir response lag was quantified as 8 days, and seismic influence was effectively represented by a short-term memory variable with a 5-day effective window and event-dependent attenuation. Incorporating these elements into an ARDL formulation improved prediction accuracy relative to a no-lag baseline and enabled a practical graded warning scheme. The Trend Strength ratio provided a physically interpretable bridge from numerical prediction to state diagnosis, supporting earlier recognition of acceleration episodes and reducing sensitivity to short-lived anomalies.

Non-line-of-sight (NLOS) propagation, multipath effects, and unstructured abrupt anomalies severely degrade the positioning accuracy of Ultra-Wideband (UWB) systems in complex indoor environments. To address this, we propose an adaptive UWB positioning method integrating variational Bayesian (VB) inference and historical residual analysis. The core contributions are threefold: (1) A Gaussian-Cauchy-Uniform mixture noise model is constructed to characterize Line-of-sight (LOS) random errors, NLOS heavy-tailed errors, and unstructured abrupt anomalies. (2) A historical residual-based normality decision mechanism extracts four statistical features from a sliding window and uses logistic regression to compute measurement reliability, providing reliable priors for VB inference. (3) A VB-Levenberg- Marquardt (VB-LM) joint optimization framework adaptively estimates noise parameters via VB inference and solves the weighted least-squares positioning problem efficiently. Extensive experiments demonstrate that the proposed method significantly outperforms mainstream algorithms, achieving superior accuracy and robustness in challenging occluded indoor scenarios.

Traditional Earthquake Early Warning (EEW) systems provide initial prediction from empirical relationship between P-wave and final magnitude, which often saturate during large events (M-W > 7). High-rate GNSS data address this limitation by directly capturing co-seismic displacement, but struggle with unilateral ruptures or sparse earthquake records. In contrast, deep learning presents an end-to-end solution for this challenging problem, with its powerful nonlinear fitting capability. In this study, we develop a real-time magnitude estimation approach for large earthquakes, which couples high-rate GNSS data with Temporal Fusion Transformer (TFT). Based on the Japan Trench structure, we simulate over 50000 earthquakes (M-W 7.0 similar to 9.5) and generate displacement waveforms across multiple GNSS stations. The end-to-end model uses raw three-component waveforms as input and outputs real-time magnitude series, while quantifying the uncertainty by quantile loss function. In simulated earthquakes of testing dataset, the deep learning model achieves over 93% accuracy within 60 s after P-wave arrival. Even with limited station availability, the model maintains similar to 80% accuracy, significantly outperforming the Peak Ground Displacement (PGD) scaling law, which delivers only similar to 70%. In real earthquake cases, the deep learning model estimates the final magnitude within 90 s, whereas traditional method shows a notable delay, requiring at least 120 s for a reliable alert during the M-W 9.1 event. The results demonstrate that deep learning can effectively extract critical information from raw waveforms, which can offer substantial improvement to current EEW systems.

The global navigation satellite system (GNSS) is essential for timing and positioning. In conventional receivers, clock offset is treated as a common error and often lacks careful modeling. However, accurate clock state estimation is crucial in GNSS-based remote timing. Current methods typically model clock error as white noise, which can amplify estimation noise in both the up-coordinate and clock states under certain conditions. Incorporating clock modeling has the potential to mitigate such noise. This study explores the theoretical foundations of clock modeling and examines its influence on GNSS positioning and timing performance. We establish the GNSS timing model and the clock signal model, and clarify the relationship between Allan Variance and the diffusion coefficient. Using a small Rubidium atomic clock and an oven controlled crystal oscillator (OCXO) as examples, we evaluate the effect of clock modeling on frequency offset estimation noise and vertical positioning precision. Theoretical and experimental results demonstrate that clock modeling significantly reduces frequency offset estimation noise, with noise attenuation ranging from 17.19% to 52.83% for OCXO and 87.67% to 97.83% for the Rubidium clock. More stable clocks exhibit greater improvement. Additionally, clock modeling enhances short-term up-coordinate positioning stability, showing improvements of 78.74% for OCXO and 84.23% for the Rubidium clock at 1 s intervals. These findings highlight the potential of clock modeling for rapid online frequency monitoring and improved GNSS timing and positioning performance with OCXOs and compact atomic clocks.

High-precision positioning plays a pivotal role in intelligent transportation systems by enabling reliable navigation, automated intersection management and high-precision platooning for autonomous vehicles in large-scale, unfamiliar environments. However, existing loosely or tightly coupled integration solutions of Global Navigation Satellite System (GNSS), Inertial Navigation System (INS), and Visual System can be challenged by the limited availability of GNSS signals, which will affect their ability to consistently achieve robust high-precision positioning in harsh environments. In this work, we propose a multi-sensor (GNSS/INS/Vision) ultra-tightly coupled system with GNSS carrier phase long coherent integration tracking (MUT-LCI). At the signal processing level, MUT-LCI utilizes the multi-sensor sensed dynamics to assist GNSS signal tracking loops and realizes extended coherent integration periods up to 300 milliseconds, improving the availability of GNSS carrier phase observations in highly challenging environments significantly. The data fusion of the GNSS RTK/INS/Vision is conducted using a Multi-state Constraint Kalman Filter (MSCKF). To evaluate the performance of the proposed MUT-LCI system, we conducted experiments using an intelligent transportation wheeled robot platform in harsh environments, including areas covered by dense tree canopies and surrounded by tall buildings. The results demonstrate that the proposed MUT-LCI system achieves precise GNSS carrier phase signal tracking in highly challenging environments, enabling continuous RTK fixed solutions and providing reliable centimeter-level positioning accuracy.

Realizing a Terrestrial Reference Frame (TRF) with an accuracy of 1 mm and a long-term stability of 0.1 mm/yr is a longstanding goal of the geodesy field. To achieve this, selecting an appropriate stochastic model to accurately characterize the nonlinear coordinate variations of geodetic stations is essential for TRF realization. However, the commonly used Random Walk (RW) model in filtering is not the optimal noise model for time-correlated noise in Global Navigation Satellite System (GNSS) coordinates. In this study, we replace the RW model with a first-order autoregressive (AR[1]) process to model the GNSS time-correlated noise and implement a GNSS TRF solution aligned with ITRF2020 via the Square Root Information Filter (SRIF). We found that the AR[1] process used in this study has a higher cut-off frequency than the RW model, allowing it to retain a larger portion of the input flicker noise. Consequently, the GNSS time-correlated noise modelled by AR[1] more closely approximates true flicker noise than that modelled by RW. When time-correlated noise is modelled by AR[1], the median RMS of coordinate residuals is decreases to 0.3 and 2.0 mm in the horizontal and up components, respectively. Moreover, the AR[1] process can capture short-term correlations in time-correlated noise parameters, thereby enhancing the accuracy of short-term (approximately 11 weeks) TRF coordinate predictions. These findings demonstrate the potential of incorporating time-correlated noise using AR[1] in GNSS data assimilation, with implications for both multi-technique global TRF realization and regional GNSS TRF solutions.

As a vital atmospheric component, water vapor influences climate change and drives numerous weather-related processes. The Global Positioning System (GPS) provides a valuable tool for investigating Precipitable Water Vapor (PWV). This study reanalyzed the raw GPS data from 2000 to 2020 at 25 stations over China to generate a consistent and reliable GPS PWV data set. To obtain homogenized GPS PWV time series, potential artificial shifts in the mean caused by changes in instruments and environmental conditions were identified and corrected based on GPS station log files and comparisons with numerical weather prediction reanalysis PWV from the ECMWF Reanalysis v5 (ERA5) and satellite PWV observations from the Moderate Resolution Imaging Spectrometer (MODIS). A comprehensive assessment of the spatiotemporal variability of PWV is performed using this homogenized GPS data set, while using the ERA5 and the MODIS PWV products for comparison. We first found that both ERA5 and MODIS are slightly wetter than GPS PWV, and ERA5 agrees better with GPS PWV than MODIS does. The annual PWV trends are generally significantly positive across China, and exhibit a west-to-east gradient in increasing PWV. The homogenized GPS PWV shows a wetting in eastern and central China during summer and a drying or no trend in winter over central and northern China. The relationships among PWV, specific humidity, and precipitation with temperature more closely follow the scale of the Clausius-Clapeyron expectation in monsoon-dominated regions than in the plateau region. This study provides new insights into China's atmospheric moisture dynamics in a changing climate.

To address the performance degradation of vehicle-borne integrated navigation caused by frequent GNSS signal attenuation in complex urban environments, and to overcome the limitations of conventional non-holonomic constraints (NHC) and zero-velocity updates (ZUPT), which rely on static empirical noise parameters, fail to adapt to time-varying vehicle dynamics, and are prone to inducing filter divergence, this paper proposes a Global Navigation Satellite System (GNSS) Real-Time Kinematic (RTK)/Inertial Navigation System (INS)/odometer (ODO) integrated navigation algorithm that fuses Bi-directional Long Short-Term Memory (BiLSTM)-based motion state recognition with adaptive NHC/ZUPT noise tuning. First, a lightweight temporal classifier based on BiLSTM is constructed. By extracting temporal dependencies from raw multi-sensor observations, the classifier achieves real-time, high-accuracy recognition of vehicle stationary, straight-line, and turning states, overcoming the misclassification problem of conventional threshold-based methods in state transition regions. Second, a dynamic noise tuning mechanism driven by state probabilities is designed: under stationary conditions, a polynomial decay strategy is employed to tighten the ZUPT noise covariance, thereby suppressing the divergence of inertial sensor biases; under turning conditions, an exponential expansion strategy is adopted to dynamically inflate the NHC noise covariance, relaxing the lateral velocity constraint that would otherwise violate the Ackermann steering assumption. Finally, the adaptive measurement noise matrix is incorporated into a Kalman filtering framework for multi-source information fusion. Real-world vehicle tests demonstrate that, compared with conventional fixed-parameter filtering methods, the proposed method effectively suppresses filter oscillations caused by state misclassification, enhances positioning robustness and resilience under both stationary and turning conditions, and ultimately improves positioning accuracy.

Aiming at the problem that the pseudo-velocity measurement noise of non-holonomic constraints (NHCs) in the integrated navigation of vehicle-mounted a global navigation satellite system/inertial navigation system (GNSS/INS) is time-varying and thick-tailed in complex road conditions (turning, sideslip, etc.) and cannot be accurately predicted, an adaptive estimation method for the initial value of NHC lateral velocity noise based on multiple linear regression is proposed. On the basis of this method, a Gaussian Student’s T distribution variational Bayesian filtering algorithm (Ga-St VBAKF) based on NHC pseudo-velocity measurement noise modeling is proposed through modeling and analysis of pseudo-velocity measurement noise. Firstly, in order to adaptively adjust the initial value of NHC lateral velocity noise, a vehicle turning detection algorithm is used to detect whether the vehicle is turning. Secondly, based on the vehicle motion state, the variational Bayesian method is used to adaptively estimate the statistical characteristics of the measurement noise in real time based on modeling of the lateral velocity noise as Gaussian white noise or Student’s T distribution thick-tail noise. The test results show that compared to the traditional Kalman filtering algorithm with fixed noise, the Ga-St VBAKF algorithm with noise adaptation reduces the maximum horizontal position error by 65.9% in the GNSS/NHC/OD/INS (where OD stands for odometer and INS stands for inertial measurement unit) system when the vehicle is in a turning state, and by 42.3% in the NHC/OD/INS system. This indicates that the algorithm can effectively suppress the divergence of positioning errors during turning and improve the performance of integrated navigation.

Human perception possesses the remarkable ability to mentally reconstruct the complete structure of occluded objects, which has inspired researchers to pursue amodal instance segmentation for a more comprehensive understanding of the scene. Previous works have shown promising results, but they often capture the contextual dependencies in an unsupervised way, which can lead to undesirable contextual dependencies and unreasonable feature representations. To tackle this problem, we propose a Pixel Affinity-Parsing (PAP) module trained with the Pixel Affinity Loss (PAL). Embedded into CNN, the PAP module can leverage learned contextual priors to guide the network to explicitly distinguish different relationships between pixels, thus capturing the intra-class and inter-class contextual dependencies in a non-local and supervised way. This process helps to yield robust feature representations to prevent the network from misjudging. To demonstrate the effectiveness of the PAP module, we design an effective Pixel Affinity-Parsing Network (PAPNet). Notably, PAPNet also introduces shape priors to guide the amodal mask refinement process, thus preventing implausible shapes in the predicted masks. Consequently, with the dual guidance of contextual and shape priors, PAPNet can reconstruct the full shape of occluded objects accurately and reasonably. Experimental results demonstrate that the proposed PAPNet outperforms existing state-of-the-art methods on multiple amodal datasets. Specifically, on the KINS dataset, PAPNet achieves 37.1% AP, 60.6% AP50 and 39.8% AP75, surpassing C2F-Seg by 0.6%, 2.4% and 2.8%. On the D2SA dataset, PAPNet achieves 71.70% AP, 85.98% AP50 and 77.10% AP75, surpassing PGExp by 0.75% and 0.33% in AP50 and AP75, and being comparable to PGExp in AP. On the COCOA-cls dataset, PAPNet achieves 41.29% AP, 60.95% AP50 and 46.17% AP75, surpassing PGExp by 3.74%, 3.21% and 4.76%. On the CWALT dataset, PAPNet achieves 72.51% AP, 85.02% AP50 and 80.47% AP75, surpassing VRSPNet by 5.38%, 0.07% and 5.35%. The code is available at https://github.com/jiaoZ7688/PAP-Net.

The nonholonomic constraint (NHC) serves as a key enhancement in integrated vehicular positioning systems that center on global navigation satellite system (GNSS), inertial navigation system (INS), and odometer (ODO). Conventional fixed NHC noise covariance matrices often fail to match actual vehicle dynamics during special maneuvers (e.g., aggressive turns and skidding), while time-varying heavy-tailed lateral noise characteristics present significant prediction challenges. Existing neural network approaches, while capable of establishing inertial measurement unit (IMU)-NHC noise mappings, suffer from two key limitations: 1) insufficient robustness under complex motion states and 2) degraded accuracy or even filter divergence due to incomplete noise modeling. To address these issues, this article proposes a novel gated recurrent unit (GRU)/bidirectional long short-term memory (Bi-LSTM)-based motion state classification method for NHC noise estimation, integrated with a variational Bayesian (VB) filtering algorithm utilizing a Weibull-Student's t-distribution noise model. Two sets of GNSS/INS/ODO positioning experiments (T01/T02) in urban environments demonstrate that the proposed method significantly outperforms conventional schemes: achieving superior 3-D root-mean-squared (rms) performance of (0.36, 0.32, 0.52) m for the T01 dataset and (0.21, 0.22, 0.13) m for the T02 dataset, thereby validating both the method's effectiveness and the necessity of its constituent modules.

Attitude information serves as a critical parameter for the safe operation of autonomous platforms such as autonomous ground vehicles (AGVs). To address the challenge of continuous and reliable attitude acquisition in complex environments, this article proposes a resilient dual-antenna global navigation satellite system (GNSS)/Inertial navigation system (INS) fusion method. An error-quaternion state model is constructed within a Kalman filter framework, while an observation model integrates dual-antenna GNSS-derived attitude angles and accelerometer outputs. To resolve the critical issue of dynamic baseline ambiguity resolution, a constraint-augmented rapid ambiguity fixing method is developed. Motion-induced line-of-sight acceleration and Earth-referenced centripetal acceleration are compensated to achieve precise accelerometer-based attitude determination in dynamic scenarios. An innovative observation model incorporating angular misalignment estimation is introduced to effectively mitigate systematic errors caused by GNSS/INS axis misalignment. Experiments conducted on two typical ground platforms in GNSS-challenged environments demonstrate that both platforms-a passenger car (T01) with a dual-antenna baseline of 1.115 m and an autonomous vehicle (T02) with a baseline of 0.804 m-achieve pitch and yaw estimation accuracy better than 0.7 degrees (T01: 0.67 degrees/0.54 degrees; T02: 0.49 degrees/0.48 degrees) using dual-antenna GNSS. The GNSS/INS integrated solution further enhances performance, achieving roll/pitch/yaw accuracies of 1.05 degrees/0.32 degrees/0.19 degrees for T01 degrees and 0.38 degrees/0.16 degrees/0.25 degrees for T02, which fully validates the robustness of the proposed algorithm in complex environments.

The India-Eurasia collision zone is the largest deforming region on the planet with numerous faults and widespread earthquakes, extending from the Himalayan Front to north of the Tien Shan. Developed from plate tectonic theory, block models have long been used to describe the crustal deformation in the collision zone, and GPS data are often invoked to constrain and test the models. Although previous block models perform well against GPS data on the whole, the detailed performance in many areas of the collision zone remains uncertain due to sparsity of GPS data and the low resolution of the fault database used to define the blocks. In this study, we process the raw GPS data collected via regional continuous GPS observation networks and Crustal Movement Observation Network of China (CMONOC) up to 2021, mainly located in Tibet, and obtain our core GPS velocity field with 420 continuous and 872 campaign stations. We further incorporate published GPS velocities, mainly located in the Himalaya and Tien Shan regions. We convert these velocities into our core solution to keep all the velocities in a consistent reference frame. As a result, we provide the densest and up-to-date GPS velocity field in the India-Eurasia collision zone including 2811 stations. Although the stations from CMONOC have been presented before, our updated velocities are more robust as they are derived from a longer time span, e.g., 5 years more than Wang and Shen [2020]. Also, we add an extra 351 stations for the collision zone compared to Wang and Shen [2020], most of which are continuous stations, over 300 of which have never been published. Wright et al. [2023] presented the first high-resolution InSAR velocity field for whole Tibet. Constraints from the InSAR data enable us to effectively evaluate the detailed performance of block modeling in Tibet, especially in the remote regions where the GPS data are sparse. We incorporate the GPS and InSAR velocity fields, and 170 Quaternary fault slip rates into a recently-developed high-resolution block model with 237 blocks by Styron [2022] to predict block motion and fault slip rates throughout the collision zone. The block model fits the data well in general, although there are some significant residuals. The predicted slip rates along ~900 faults from the model are generally small except for those along several major faults, including the major Tibetan strike-slip faults, which have larger slip rates but still within the level of 10 mm/yr, and the Main Himalayan Thrust, which has a convergence rate at the level of about 15 mm/yr. The predicted slip rates show along-strike variations, and are consistent with previous geodetic studies. We then use our results to assess the limitations of tectonic block modelling for applications in seismic hazard assessment and in understanding the geodynamics of continental tectonics. The results suggest that tectonic strain has two modes: a few major faults exhibit focused strain and high slip rates; between these major structures, deformation is more continuous.