Selected work

Representative Papers

Energy-Efficient Fast Object Detection on Edge Devices for IoT Systems

Jun 01, 2025IEEE Internet of Things Journal

This work proposes an efficient object detection framework tailored for high-speed moving objects in IoT systems, addressing the critical trade-off among accuracy, latency, and energy efficiency. By integrating frame differencing with a lightweight neural network architecture that combines MobileNet, YOLOX, and Transformer components, the method achieves real-time performance while maintaining high precision. The framework is deployed and evaluated on edge devices including the AMD Alveo U50, Jetson Orin Nano, and Hailo-8. Experimental results demonstrate that, compared to conventional end-to-end approaches, the proposed solution improves average precision by 28.3%, enhances energy efficiency by 3.6×, and reduces latency by 39.3%, thereby achieving a superior balance among accuracy, energy consumption, and real-time responsiveness in high-speed scenarios.

8 citationsRead paper

Structure-Preserving Visualization of Complex Systems through Discrete Approximation: An Application to Argo Data

Aug 14, 2026

This study addresses the challenges of detail loss and inadequate representation of spatial variability in visualizing the vertical structure of complex systems. We propose a structure-preserving discrete approximation framework that integrates clustering analysis with interpretable geometric feature color encoding to effectively balance fine-scale profile structures with macroscopic spatial distributions. Leveraging million-scale Argo data, this research successfully identifies representative profile morphologies and generates a comprehensive global visualization atlas of mesopelagic temperature-salinity vertical structures. The resulting atlas simultaneously captures micro-level details and large-scale spatial variability, establishing a novel paradigm for structural analysis of complex oceanographic data.

0 citationsRead paper

Effective and Low-cost Lane-based Map Localization for Vehicle-Centric Route Generation

Jun 14, 2026

This work addresses the challenge of generating high-precision driving trajectories aligned with the driver’s perspective under low-cost hardware constraints. The authors propose OLRA, a novel framework that, for the first time, integrates map-based navigation paths with visual lane perception. By leveraging a map–vision path matching algorithm and a lightweight sensor fusion strategy, OLRA achieves consistent alignment between vehicle localization and the driver’s intuitive viewpoint. The study further introduces an application-oriented route evaluation metric suite and validates the approach on the nuScenes dataset. Results demonstrate that OLRA significantly outperforms OpenPilot in complex road segments and at distances beyond 20 meters, achieving lower overall Euclidean error and thereby enhancing the intuitiveness and reliability of driving guidance systems.

0 citationsRead paper

Focused Weighted-Average Least Squares Estimator

Mar 03, 2026

This study addresses the high computational complexity in focused model averaging arising from the exponential number of candidate submodels by proposing a computationally efficient, near-optimal Focused Weighted Least Squares (FWALS) estimator. The method introduces semi-orthogonalized auxiliary regressors to reduce the weight optimization problem to a regression scale proportional only to the number of auxiliary variables. It is solved using local zero-neighborhood asymptotic analysis, a plug-in AMSE criterion, and the Focused Information Criterion (FIC). Both theoretical analysis and simulation studies demonstrate that FWALS achieves stable performance—closely approximating the FIC benchmark—for focused targets such as impulse response functions, while substantially improving computational efficiency.

0 citationsRead paper

Timely Information Updating for Mobile Devices Without and With ML Advice

Dec 19, 2025

Mobile devices face a fundamental trade-off between timeliness and energy consumption in status updates. This paper addresses online decision-making under uncertainty—without prior knowledge—and proposes an adaptive update algorithm that integrates unreliable machine learning (ML) advice. Our method introduces a threshold-based trust mechanism for ML recommendations in adversarial settings, rigorously proving that partial trust degrades robustness. Grounded in a consistency–robustness theoretical framework, we derive the optimal competitive ratio, which scales linearly with the range of update costs. The algorithm achieves theoretical optimality under both adversarial and stochastic input models. Extensive simulations demonstrate its significant superiority over baseline approaches and strong robustness against ML prediction noise and multi-source uncertainties.

0 citationsRead paper

Recent publications

Latest Papers

Structure-Preserving Visualization of Complex Systems through Discrete Approximation: An Application to Argo Data

Aug 14, 2026

This study addresses the challenges of detail loss and inadequate representation of spatial variability in visualizing the vertical structure of complex systems. We propose a structure-preserving discrete approximation framework that integrates clustering analysis with interpretable geometric feature color encoding to effectively balance fine-scale profile structures with macroscopic spatial distributions. Leveraging million-scale Argo data, this research successfully identifies representative profile morphologies and generates a comprehensive global visualization atlas of mesopelagic temperature-salinity vertical structures. The resulting atlas simultaneously captures micro-level details and large-scale spatial variability, establishing a novel paradigm for structural analysis of complex oceanographic data.

0 citationsRead paper

Effective and Low-cost Lane-based Map Localization for Vehicle-Centric Route Generation

Jun 14, 2026

This work addresses the challenge of generating high-precision driving trajectories aligned with the driver’s perspective under low-cost hardware constraints. The authors propose OLRA, a novel framework that, for the first time, integrates map-based navigation paths with visual lane perception. By leveraging a map–vision path matching algorithm and a lightweight sensor fusion strategy, OLRA achieves consistent alignment between vehicle localization and the driver’s intuitive viewpoint. The study further introduces an application-oriented route evaluation metric suite and validates the approach on the nuScenes dataset. Results demonstrate that OLRA significantly outperforms OpenPilot in complex road segments and at distances beyond 20 meters, achieving lower overall Euclidean error and thereby enhancing the intuitiveness and reliability of driving guidance systems.

0 citationsRead paper

Focused Weighted-Average Least Squares Estimator

Mar 03, 2026

This study addresses the high computational complexity in focused model averaging arising from the exponential number of candidate submodels by proposing a computationally efficient, near-optimal Focused Weighted Least Squares (FWALS) estimator. The method introduces semi-orthogonalized auxiliary regressors to reduce the weight optimization problem to a regression scale proportional only to the number of auxiliary variables. It is solved using local zero-neighborhood asymptotic analysis, a plug-in AMSE criterion, and the Focused Information Criterion (FIC). Both theoretical analysis and simulation studies demonstrate that FWALS achieves stable performance—closely approximating the FIC benchmark—for focused targets such as impulse response functions, while substantially improving computational efficiency.

0 citationsRead paper

Timely Information Updating for Mobile Devices Without and With ML Advice

Dec 19, 2025

Mobile devices face a fundamental trade-off between timeliness and energy consumption in status updates. This paper addresses online decision-making under uncertainty—without prior knowledge—and proposes an adaptive update algorithm that integrates unreliable machine learning (ML) advice. Our method introduces a threshold-based trust mechanism for ML recommendations in adversarial settings, rigorously proving that partial trust degrades robustness. Grounded in a consistency–robustness theoretical framework, we derive the optimal competitive ratio, which scales linearly with the range of update costs. The algorithm achieves theoretical optimality under both adversarial and stochastic input models. Extensive simulations demonstrate its significant superiority over baseline approaches and strong robustness against ML prediction noise and multi-source uncertainties.

0 citationsRead paper

An Accurate Standard Error Estimation for Quadratic Exponential Logistic Regressions by Applying Generalized Estimating Equations to Pseudo-Likelihoods

Sep 30, 2025

In quadratic exponential binary distribution (QIBD) regression models, standard errors estimated via pseudolikelihood are severely underestimated. To address this, we propose a novel standard error correction method that integrates pseudolikelihood estimation with generalized estimating equations (GEE). Theoretically, we prove that adopting an independence working correlation structure within the GEE framework ensures consistent parameter estimation, whereas misspecifying the dependence structure induces substantial bias. Through analytical derivation and extensive simulations across diverse dependency scenarios, our method demonstrably improves both accuracy and robustness of standard error estimation. Empirical applications to toxicological longitudinal data and constitutional court judgment network data confirm its strong performance under realistic, complex dependency structures. This work provides the first standard error estimator for QIBD-type models that simultaneously achieves computational efficiency and statistical reliability.

0 citationsRead paper