Selected work

Representative Papers

Improving Remote Patient Monitoring Systems Using a Fog-Based IoT Platform With Speech Recognition

Aug 01, 2023IEEE Sensors Journal

To address network congestion, privacy breaches, and inefficient human–machine interaction arising from data explosion in Remote Patient Monitoring (RPM), this paper proposes a fog-enhanced IoT-based RPM architecture integrated with on-device speech recognition. The system performs real-time sensor data processing and localized privacy-preserving operations at the edge, substantially reducing cloud workload; concurrently, it incorporates a lightweight speech recognition module to enable natural-language-driven clinician–patient interaction. Its key innovation lies in the first synergistic integration of fog computing and edge-side speech understanding within RPM systems, enabling resource-adaptive scheduling and ultra-low-latency response. Experimental results demonstrate an average end-to-end latency of <120 ms, a 3.2× throughput improvement over baseline approaches, and a 96.7% accuracy in voice command recognition—validating the framework’s superior performance in real-time responsiveness, data security, and interactive usability.

8 citationsRead paper

Boundary-Enhanced Segmentation of Pig Point Clouds in Commercial Housing Environments

Aug 12, 2026

This study addresses the challenges of ambiguous boundaries, local adhesion, and erroneous segmentation between pigs and background in commercial pig barns by proposing a boundary-aware point cloud segmentation method. The approach employs an Octree Transformer backbone to effectively integrate fine-grained local geometric details with global semantic context. It introduces soft-distance boundary pseudo-labels for continuous boundary supervision and incorporates a novel bidirectional cross-boundary semantic module to explicitly model interactions between boundary cues and semantic features. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art models on a comprehensive dataset, achieving superior performance in segmentation accuracy, mean Intersection over Union (mIoU), and boundary delineation, thereby providing high-quality point cloud inputs for precision livestock farming.

0 citationsRead paper

AI Forensics Across White-, Grey-, and Black-Box Access: A Process Model and Research Agenda for Post-Incident Investigation of AI Systems

Aug 04, 2026

This study addresses the absence of a unified investigative framework for AI system incidents, which hinders behavior reconstruction, accountability attribution, and supply chain traceability. It proposes the first AI forensics process model structured around investigators’ levels of system access—white-box, gray-box, and black-box—and organized into four phases: collection, preservation, analysis, and reporting. The work introduces a method for ranking evidence volatility and integrates multi-source data—including logs, context windows, retrieval corpora, and training lineages—into a structured forensic workflow matrix. By clarifying critical open issues such as black-box evidence preservation and model version authentication, the paper identifies four core challenges and establishes a theoretical foundation for designing auditable and accountable AI systems.

0 citationsRead paper

Future Rendering $\neq$ Future Surface: A Benchmark and Dataset for Dynamic Surface Reconstruction Beyond the Observed Window

Jul 23, 2026

Current dynamic scene reconstruction methods lack standardized evaluation protocols for future geometric surfaces (“future surfaces”), hindering their deployment in temporally sensitive applications such as augmented reality and robotics. This work proposes FutureSurf—the first diagnostic benchmark specifically designed for future surface reconstruction—leveraging analytically controlled motion sequences to generate precise per-frame mesh ground truth, with training on the first 75% of the temporal sequence and evaluation on the remaining 25%. Using DG-Mesh and Deformable-3DGS backbones combined with deformation MLPs for temporal modeling and Chamfer distance for scoring, experiments reveal that existing methods exhibit 2.7–6.6× higher error on future frames compared to observed frames. The proposed falsifiable control mechanism proves effective, with future errors predominantly localized in moving regions. Notably, no statistically significant correlation is found between future rendering quality and geometric accuracy.

0 citationsRead paper

Conditional copula graphic estimator for semi-competing risks data

Jul 10, 2026

This study addresses the challenge of analyzing non-terminal event times in semi-competing risks data, where dependent censoring induced by terminal events and unadjusted covariates often leads to confounding bias. The authors propose a conditional Copula graphical estimation method that simultaneously incorporates covariates into both the marginal survival functions and the Copula dependence structure, enabling fully conditional modeling. Within a semiparametric framework, the approach employs Archimedean Copulas to capture covariate-dependent associations, estimates marginal distributions and dependence parameters nonparametrically, and solves the resulting system via an alternating iterative algorithm. Simulation studies and real-data analyses demonstrate that, compared to unconditional methods, the proposed approach substantially improves estimation accuracy, underscoring the critical role of covariate adjustment in semi-competing risks analysis.

0 citationsRead paper

Recent publications

Latest Papers

Boundary-Enhanced Segmentation of Pig Point Clouds in Commercial Housing Environments

Aug 12, 2026

This study addresses the challenges of ambiguous boundaries, local adhesion, and erroneous segmentation between pigs and background in commercial pig barns by proposing a boundary-aware point cloud segmentation method. The approach employs an Octree Transformer backbone to effectively integrate fine-grained local geometric details with global semantic context. It introduces soft-distance boundary pseudo-labels for continuous boundary supervision and incorporates a novel bidirectional cross-boundary semantic module to explicitly model interactions between boundary cues and semantic features. Experimental results demonstrate that the proposed method significantly outperforms state-of-the-art models on a comprehensive dataset, achieving superior performance in segmentation accuracy, mean Intersection over Union (mIoU), and boundary delineation, thereby providing high-quality point cloud inputs for precision livestock farming.

0 citationsRead paper

AI Forensics Across White-, Grey-, and Black-Box Access: A Process Model and Research Agenda for Post-Incident Investigation of AI Systems

Aug 04, 2026

This study addresses the absence of a unified investigative framework for AI system incidents, which hinders behavior reconstruction, accountability attribution, and supply chain traceability. It proposes the first AI forensics process model structured around investigators’ levels of system access—white-box, gray-box, and black-box—and organized into four phases: collection, preservation, analysis, and reporting. The work introduces a method for ranking evidence volatility and integrates multi-source data—including logs, context windows, retrieval corpora, and training lineages—into a structured forensic workflow matrix. By clarifying critical open issues such as black-box evidence preservation and model version authentication, the paper identifies four core challenges and establishes a theoretical foundation for designing auditable and accountable AI systems.

0 citationsRead paper

Future Rendering $\neq$ Future Surface: A Benchmark and Dataset for Dynamic Surface Reconstruction Beyond the Observed Window

Jul 23, 2026

Current dynamic scene reconstruction methods lack standardized evaluation protocols for future geometric surfaces (“future surfaces”), hindering their deployment in temporally sensitive applications such as augmented reality and robotics. This work proposes FutureSurf—the first diagnostic benchmark specifically designed for future surface reconstruction—leveraging analytically controlled motion sequences to generate precise per-frame mesh ground truth, with training on the first 75% of the temporal sequence and evaluation on the remaining 25%. Using DG-Mesh and Deformable-3DGS backbones combined with deformation MLPs for temporal modeling and Chamfer distance for scoring, experiments reveal that existing methods exhibit 2.7–6.6× higher error on future frames compared to observed frames. The proposed falsifiable control mechanism proves effective, with future errors predominantly localized in moving regions. Notably, no statistically significant correlation is found between future rendering quality and geometric accuracy.

0 citationsRead paper

Conditional copula graphic estimator for semi-competing risks data

Jul 10, 2026

This study addresses the challenge of analyzing non-terminal event times in semi-competing risks data, where dependent censoring induced by terminal events and unadjusted covariates often leads to confounding bias. The authors propose a conditional Copula graphical estimation method that simultaneously incorporates covariates into both the marginal survival functions and the Copula dependence structure, enabling fully conditional modeling. Within a semiparametric framework, the approach employs Archimedean Copulas to capture covariate-dependent associations, estimates marginal distributions and dependence parameters nonparametrically, and solves the resulting system via an alternating iterative algorithm. Simulation studies and real-data analyses demonstrate that, compared to unconditional methods, the proposed approach substantially improves estimation accuracy, underscoring the critical role of covariate adjustment in semi-competing risks analysis.

0 citationsRead paper

HSA: Hierarchical Slot Attention for Multi-granularity Scene-Decomposition

Jul 09, 2026

Existing slot attention methods support only single-granularity, appearance-driven scene decomposition, making it difficult to model semantic hierarchies such as foreground/background, categories, and instances. This work proposes a hierarchical slot attention mechanism that enables end-to-end joint learning of scene decomposition at three granularities—holistic, semantic, and panoptic—within a single model. The approach requires only 10% of labeled data and a hierarchical alignment loss for effective training. It is the first method to achieve multi-granular semantic decomposition without relying on multiple models and introduces novel metrics—group purity and inclusiveness—to validate hierarchical structure encoding in the representation space. On COCO and PASCAL VOC, the method substantially outperforms the strongest single-granularity baselines, improving Adjusted Rand Index (ARI) by 41.5, 14.6, and 10.4 at holistic, semantic, and panoptic levels on COCO, with even greater gains observed on PASCAL VOC.

0 citationsRead paper