26 September 2022 in IEEE Sensors Journal
Accurate and robust localization is a fundamental need for mobile agents. Visual-inertial odometry (VIO) algorithms exploit the information from camera and inertial sensors to estimate position and translation. Recent deep learning based VIO models attract attentions as they provide pose information in a data-driven way, without the need of designing hand-crafted algorithms. Existing learning based VIO models rely on recurrent models to fuse multimodal data and process sensor signal, which are hard to train and not efficient enough. We propose a novel learning based VIO framework with external memory attention that effectively and efficiently combines visual and inertial features for states estimation. Our proposed model is able to estimate pose accurately and robustly, even in challenging scenarios, e.g., on overcast days and water-filled ground, which are difficult for traditional VIO algorithms to extract visual features. Experiments validate that it outperforms both traditional and learning based VIO baselines in different scenes.
11 December 2025 in IEEE Transactions on Instrumentation and Measurement
Large-scale camera localization in driving environments is pivotal in autonomous driving and robotics, where deep learning-based solutions offer promise for end-to-end problem-solving. However, existing learning-based camera pose regression methods often encounter accuracy and robustness challenges in complex environments, hampering scalability to large scenarios and datasets. This work introduces EMaxLoc, a novel end-to-end camera pose regression model that integrates a Multi-axis Vision Transformer (MaxViT) with an Efficient Multi-scale Attention (EMA) module for large-scale localization. The introduced MaxViT extractor captures both local and global spatial features through its block and grid attention mechanisms, while the EMA module enhances multi-scale feature fusion and channel-wise contextual modeling. EMaxLoc demonstrates remarkable proficiency in environments characterized by scene dynamics and illumination changes. Furthermore, our Transformer-based framework showcases scalability, suggesting potential for continual improvement in localization performance. Extensive experiments validate EMaxLoc’s superior performance with increasing dataset sizes, surpassing existing deep learning-based methods such as AtLoc, PoseNet, MapNet and EffLoc in large-scale driving environments.
2 April 2026 in IEEE Transactions on Instrumentation and Measurement
Dual-foot pedestrian navigation systems leveraging zero-velocity updates (ZUPTs) can effectively reduce heading drift by incorporating inter-foot ranging constraints. However, their performance is highly sensitive to ranging noise and zero-velocity detector accuracy–misdetections introduce outlier observations that degrade state estimation. Conventional robust kernel-based outlier suppression mitigates this issue but introduces strong non-convexities, making the optimization prone to convergence at local minima. To address these challenges, we propose a dual-foot Fpedestrian inertial navigation algorithm based on factor graph optimization (FGO) with graduated non-convexity (GNC). The system fuses foot-mounted IMU measurements, zero-velocity constraints, and inter-foot ranging information within a sliding-window factor graph. A GNC-based backend adaptively reweights factor contributions, suppressing the effect of outliers while progressively reducing the non-convexity introduced by robust kernels. Extensive experiments demonstrate that the proposed method enhances robustness and positioning accuracy across diverse and complex pedestrian motion conditions, outperforming conventional EKF-based and standard FGO approaches.