HZHai Zhang

Hai Zhang

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hai Zhang has a total of 42 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    A Pose-Only Point–Line Geometric Representation Framework for Visual–Inertial Localization of UAVs and UGVs in Complex Environments

    by Aiping Wang, Zhaolong Yang, Shuwen Chen, Hai Zhang

    Accurate visual–inertial localization is essential for unmanned aerial vehicles (UAVs) and unmanned ground vehicles (UGVs) operating in challenging environments. However, conventional multi-state constraint Kalman filter (MSCKF)-based visual–inertial odometry (VIO) uses landmark-based measurements with null-space projection and often delays visual updates until feature tracks end, while point-based systems may provide insufficient geometric constraints under low texture, motion blur, and illumination variations. To overcome these limitations, we extend our preliminary conference work, POPL-VIO, into a complete state estimation framework built upon pose-only (PO) representations for both point and line features. Specifically, a unified pose-only measurement model is developed for MSCKF point and line features. Feature geometry is substituted before linearization, avoiding independent landmark Jacobians and null-space projection while enabling state updates as soon as three valid views are available. Long-track SLAM points and lines are retained in the hybrid filter state and updated using landmark-based measurements. To improve the quality of geometric constraints, the unified base-frame selection algorithm of POPL-VIO is refined and analyzed for both feature types. Additionally, to ensure observational reliability in challenging environments, a stringent outlier rejection mechanism integrating image grid segmentation and bidirectional optical flow consistency is proposed. An observability analysis shows that the PO point and line measurements preserve the four fundamental unobservable directions of visual–inertial navigation, both in the linearized system and, under compatible first-estimate Jacobians, in the filter. The method is evaluated on the EuRoC MAV and KAIST Urban datasets, as well as self-collected indoor sequences acquired by manually carrying a UAV-configured sensor suite to produce UAV-like sensor motion and separate ground-platform sequences. The proposed method achieves the lowest mean ATE among the evaluated methods on EuRoC and KAIST, with mean errors of 0.103 m and 9.07 m, respectively, and also improves the mean results on the self-collected sequences. On the EuRoC MH_01 sequence, it achieves an average processing rate of approximately 35 Hz on the tested PC. These results indicate that the proposed pose-only formulation improves localization accuracy while maintaining processing efficiency comparable to that of the corresponding point-line system without the pose-only formulation.

  • Open Access

    Online Bias Estimation for Single-Platform Airborne Radar Using Bias-Subspace Information-Guided MAP-EKF

    by Junwu Luo, Xujun Guan, Chuang Song, Hai Zhang

    Systematic measurement biases are a persistent source of degradation in airborne radar target tracking. Online bias estimation is especially difficult for single-platform operations because only one measurement stream is available, and the target state and radar biases are coupled in the same nonlinear observation model. Under weakly observable geometries, directly augmenting bias states into a recursive filter may lead to slow convergence, while fixed or overly frequent batch optimization may inject updates from weakly informative windows. To address this problem, this paper proposes a Bias Subspace Information-Guided MAP-EKF (maximum a posteriori–extended Kalman filter) (BI-MAP-EKF) for online bias estimation in single-platform airborne radar. A nine-dimensional augmented state jointly describes target motion and range, azimuth, and elevation biases. A posterior Cramér–Rao lower bound (PCRLB) is constructed for this state, where the EKF prior, windowed radar measurements, and process noise propagation are jointly considered. The bias subspace is then extracted by Schur complement, and the resulting azimuth bias PCRLB is used to decide whether the MAP update is reliable enough for EKF injection. The scheduler also includes a cooldown interval and a maximum-window safeguard, which respectively limit excessive updates under informative maneuvers and prevent indefinite waiting under weak geometries. Fifty-run Monte Carlo simulations under different maneuvering conditions show that the proposed scheduler is particularly effective in weakly informative geometries, where it improves azimuth bias and horizontal position estimation while reducing the number of accepted MAP refinements compared with the tested fixed-period MAP and moving-horizon estimation (MHE) baselines. In stronger maneuvering scenarios, it provides a competitive accuracy–cost trade-off rather than uniformly outperforming aggressive MHE in every channel.

  • Open Access

    Triangle Attribute-Guided Registration for Robust Multi-Target Association in Multi-UAV Airborne Radar Sensing

    by Xujun Guan, Junwu Luo, Hai Zhang

    In multi-UAV airborne radar sensing, target-level fusion cannot start until detections reported in different UAV coordinate frames have been associated. This paper focuses on a single-epoch case: two UAV-mounted radars observe a common target group, but the reported point sets are sparse, unordered, only partially overlapping, and corrupted by angular noise, missed detections, and additional observations. In this setting, appearance cues, long tracks, and dense point-cloud neighborhoods are absent, so the association problem is mainly a sparse geometric registration problem. We propose a Triangle Attribute-Guided Robust Association (TARA) framework. TARA describes each target triple by its absolute scale and two normalized side-length ratios. These attributes are not used as final identity labels. Instead, they retrieve, verify, and rank plausible cross-view triangle pairs, from which a fixed number of rigid-transformation hypotheses are estimated. The best coarse transformation is then refined by the trimmed ICP, and the final correspondences are obtained by gated one-to-one matching in the aligned frame. The ablation results show why the attribute-guided step matters. With only 100 evaluated hypotheses, TARA reaches 93.7% F1 in the medium setting, while blind RANSAC needs about 10,000 random hypotheses to reach comparable accuracy. In composite degradation tests, TARA obtains mean F1-scores of 98.7%, 94.2%, 69.3%, and 69.1% from low to extreme conditions. The advantage is clearest when clutter, missing detections, and angular uncertainty appear together because the triangle attributes make the limited hypothesis budget much less likely to be spent on geometrically implausible triples.

  • Pseudoredundant Measurement-Based Adaptive Filtering for Estimation of Measurement Noise Covariance

    by Liuyang Jiang, Shuo Tian, Hongliang Huang, Hai Zhang

    Accurate specification of the process-noise covariance Q and measurement-noise covariance R is essential for optimal Kalman filtering. Conventional adaptive algorithms typically estimate R from innovations, thereby coupling the R update with state estimation errors and causing adaptation bias under model mismatch. This paper proposes a pseudo-measurement second-order mutual difference (PMSOMD) framework to robustly estimate R in single-measurement systems. The key idea is to construct two pseudo-measurements from state predictions at consecutive time instants and apply a first-order self-differencing (FOSD) operation to the sequence. This effectively suppresses the dominant modeling-error terms, thereby reducing the sensitivity to inaccurate Q and enabling reliable R adaptation even under process-noise mismatch. Simulation results in a maneuvering target-tracking scenario demonstrate that PMSOMD achieves more accurate and stable R estimation compared to the innovation-based adaptive estimation (IAE) and the limited memory-based random-weighted Kalman filter (LM-RWKF), particularly when the assumed Q is scaled by factors of 10 to 100. Real-world experiments further validate the effectiveness of the proposed method.

  • Open Access

    POPL-KF: A Pose-Only Geometric Representation-Based Kalman Filter for Point-Line-Based Visual-Inertial Odometry

    by Aiping Wang, Zhaolong Yang, Shuwen Chen, Hai Zhang

    Mainstream Visual-inertial odometry (VIO) systems rely on point features for motion estimation and localization. However, their performance degrades in challenging scenarios. Moreover, the localization accuracy of multi-state constraint Kalman filter (MSCKF)-based VIO systems suffers from linearization errors associated with feature 3D coordinates and delayed measurement updates. To improve the performance of VIO in challenging scenes, we first propose a pose-only geometric representation for line features. Building on this, we develop POPL-KF, a Kalman filter-based VIO system that employs a pose-only geometric representation for both point and line features. POPL-KF mitigates linearization errors by explicitly eliminating both point and line feature coordinates from the measurement equations, while enabling immediate update of visual measurements. We also design a unified base-frames selection algorithm for both point and line features to ensure optimal constraints on camera poses within the pose-only measurement model. To further improve line feature quality, a line feature filter based on image grid segmentation and bidirectional optical flow consistency is proposed. Our system is evaluated on public datasets and real-world experiments, demonstrating that POPL-KF outperforms the state-of-the-art (SOTA) filter-based methods (OpenVINS, PO-KF) and optimization-based methods (PL-VINS, EPLF-VINS), while maintaining real-time performance.

  • Adaptive unscented Kalman filter for nonlinear continuous-discrete systems based on the degree of nonlinearity

    by Aiping Wang, Hai Zhang, Xuanqi Huang, Zhaolong Yang

    In the case of strong nonlinearity and limited measurement update frequency, the precision of the UKF declines significantly because of decreases in state transition accuracy and correction ability. A smaller step size can improve estimate accuracy, but it will increase the computational load. This paper proposes an adaptive step size UKF for a nonlinear model with continuous time dynamics and discrete time measurements, which can balance precision and computation cost. A novel dynamic nonlinearity detection algorithm is proposed, based on the cosine similarity and the Euclidean distance ratio similarity of a sigma point pair after nonlinear transformation. Meanwhile, in order to avoid the increase of computation due to the nonlinearity of the non-filter performance-critical states, a sigma point construction scheme that considers filter performance-critical states has been designed. Moreover, we introduce a guideline to adaptively select appropriate step sizes based on the degree of nonlinearity in performance-critical states. In the strong nonlinear interval, the method switches to smaller step sizes and takes the obtained prediction into the update at the entire time step. Simulation results demonstrate that under highly nonlinear system conditions, and when the measurement frequency is limited, the proposed method outperforms several reference algorithms in precision and real-time performance.


  • Open Access

    An Adaptive Threshold-Based Pixel Point Tracking Algorithm Using Reference Features Leveraging the Multi-State Constrained Kalman Filter Feature Point Triangulation Technique for Depth Mapping the Environment

    by Zohaib Wahab Memon, Yu Chen, Hai Zhang

    Monocular visual–inertial odometry based on the MSCKF algorithm has demonstrated computational efficiency even with limited resources. Moreover, the MSCKF-VIO is primarily designed for localization tasks, where environmental features such as points, lines, and planes are tracked across consecutive images. These tracked features are subsequently triangulated using the historical IMU/camera poses in the state vector to perform measurement updates. Although feature points can be extracted and tracked using traditional techniques followed by the MSCKF feature point triangulation algorithm, the number of feature points in the image is often insufficient to capture the depth of the entire environment. This limitation arises from traditional feature point extraction and tracking techniques in environments with textureless planes. To address this problem, we propose an algorithm for extracting and tracking pixel points to estimate the depth of each grid in the image, which is segmented into numerous grids. When feature points cannot be extracted from a grid, any arbitrary pixel without features, preferably on the contour, can be selected as a candidate point. The combination of feature-rich and featureless pixel points is initially tracked using traditional techniques such as optical flow. When these traditional methods fail to track a given point, the proposed method utilizes the geometry of triangulated features in adjacent images as a reference for tracking. After successful tracking and triangulation, this approach results in a more detailed depth map of the environment. The proposed method has been implemented within the OpenVINS environment and tested on various open-source datasets supported by OpenVINS to validate the findings. Tracking arbitrary featureless pixel points alongside traditional features ensures a real-time depth map of the surroundings, which can be applied to various applications, including obstacle detection, collision avoidance, and path planning.


  • Open Access

    Robust Regression and Redundant Measurement Noise Estimation Adaptive Filtering for Localization in Urban Environments

    by Li Zha, Hai Zhang, Aiping Wang, Cancan Tao

    This paper focuses on a solution of target self-positioning when a Global Navigation Satellite System (GNSS) is denied. It is composed of Inertial Navigation Systems (INS), Signals of Opportunities (SOPs), and a navigation prototype. One of the options for navigation via SOP (NAVSOP) is to utilize cellular signals, such as Long Time Evolution (LTE). When the prior information is insufficient, the location of the base station (BS) is obtained by collecting the demodulation of the downlink signal, and the synchronization signal is used for static time offset correction. In view of the large positioning error of the trilateral positioning method based on Received Signal Strength (RSS), a multi-station positioning optimization method is proposed by introducing the robust regression. Monte Carlo simulation experiments indicate that the method has improved the positioning failure and insufficient accuracy. Aiming at the influence of the state estimation errors on filtering results, the Second Order Mutual Difference (SOMD) method with the noise covariance R, which is independent of the existing Extended Kalman Filter (EKF) framework and combined with Redundant Measurement Noise Covariance Estimation (RMNCE), is applied to the model. The simulation results show that the average error of the robust model is 10.28 m, which is better than the EKF method. Finally, a vehicle test in constant speed has been carried out. The results show that the proposed model can realize self-positioning with limited BS location information, and the positioning accuracy can reach 11.68 m over a 283 m trajectory.

  • Open Access

    Distributed Cognitive Positioning System Based on Nearest Neighbor Association and Multi-Point Filter Initiation for UAVs Using DTMB and INS

    by Li Zha, Hai Zhang, Na Wang, Cancan Tao, Kunfeng Lv, Ruirui Zhang

    Location is critical for the safe and effective completion of Unmanned Aerial Vehicle (UAV) missions. Since positioning errors tend to accumulate over time, uncorrected measurements from Inertial Navigation Systems (INSs) are unreliable. Aiming for UAV self-positioning under the challenges of a Global Navigation Satellite System (GNSS), this article integrates Digital Terrestrial Multimedia Broadcast (DTMB) signals and assisted INS components as external radiation sources for system design. The trigonometric geometry algorithm is proposed to estimate the pseudo-measurement, and the impact factors of the positioning error are analyzed. After filtering the pseudo-measurement by multi-point initiation, we designed a model for cross-regional positioning scenarios using the nearest-neighbor navigation association and scalar weighted distributed fusion. The simulation results demonstrate that the model can effectively track the target. Finally, the effectiveness of the positioning at a constant altitude is evaluated through different vehicle-mounted scenarios with a speed of 60 km/h. The experimental results show that the minimum positioning error can reach 18.95 m over a 525 m trajectory, thus meeting actual UAV requirements and having practical value.