Quang-Anh Nguyen-Duc
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Quang-Anh Nguyen-Duc has a total of 59 co-authors. Here are the top co-authors by number of shared publications:
- Thai Dinh Kim
Vietnam National University, Hanoi
8 shared publications - Duc Minh Pham
Vietnam National University, Hanoi
5 shared publications Minh-Anh Nguyen
Vietnam National University, Hanoi
4 shared publications- Manh-Hung Ha
Vietnam National University, Hanoi
3 shared publications Minh-Duc Pham
Computer Vision and Robotics Laboratory, International School, Vietnam National University, Hanoi, Vietnam
3 shared publications- Hai Xuan Le
International School, Vietnam National University, Hanoi, Vietnam
2 shared publications
Latest Publications
- Open Access
- 3 November 2026
- Published by MDPI AG in Preprints.org
A Comprehensive Survey of 4D Gaussian Splatting for Dynamic Scene Reconstruction
4D Gaussian splatting (4DGS) has rapidly become a prominent approach to dynamic scene reconstruction and free-viewpoint rendering, inheriting the explicit primitives and real-time rasterization of 3D Gaussian splatting while adding a temporal axis. Although 4DGS is a nascent area of research, the number of methods has grown explosively, making it difficult to navigate. In this paper, we survey recent progress in 4DGS. Specifically, we cover three major themes: (i) representations, organized into deformation-based, native 4D-primitive, hybrid/decomposition, anchor/stream- ing, and feed-forward/generalizable families; (ii) compression and storage efficiency, the defining practical bottleneck of 4DGS; (iii) applications, including autonomous driving, 4D content generation, editing and semantics, event and constrained capture, and avatars/robotics/medicine. We also present comparative results on widely used benchmarks, together with observations and future research directions. A taxonomy and chronological overview of representative methods are provided throughout this survey to help readers situate individual works.
- 3 November 2026
- Published by MDPI AG in Preprints.org
Open AccessA Comprehensive Survey of 4D Gaussian Splatting for Dynamic Scene Reconstruction
4D Gaussian splatting (4DGS) has rapidly become a prominent approach to dynamic scene reconstruction and free-viewpoint rendering, inheriting the explicit primitives and real-time rasterization of 3D Gaussian splatting while adding a temporal axis. Although 4DGS is a nascent area of research, the number of methods has grown explosively, making it difficult to navigate. In this paper, we survey recent progress in 4DGS. Specifically, we cover three major themes: (i) representations, organized into deformation-based, native 4D-primitive, hybrid/decomposition, anchor/stream- ing, and feed-forward/generalizable families; (ii) compression and storage efficiency, the defining practical bottleneck of 4DGS; (iii) applications, including autonomous driving, 4D content generation, editing and semantics, event and constrained capture, and avatars/robotics/medicine. We also present comparative results on widely used benchmarks, together with observations and future research directions. A taxonomy and chronological overview of representative methods are provided throughout this survey to help readers situate individual works.
- Open Access
CSTAR-Det: A lightweight vision-based framework for conveyor-based cashew kernel quality assessment
- 15 July 2026
- Published by Springer Nature
- p. 16-28
A Credit Risk Digital Twin Framework with Explainability-Guided Feature Engineering and Counterfactual Simulation
- 15 July 2026
- Published by Springer Nature
- p. 16-28
A Credit Risk Digital Twin Framework with Explainability-Guided Feature Engineering and Counterfactual Simulation
- Open Access
- 4 July 2026
- Published by Elsevier in SSRN Electronic Journal
AM-YOLO: Efficient Linear-Complexity Global Context Modeling for UAV Multi-Object Tracking
Multi-object tracking (MOT) from unmanned aerial vehicles (UAVs) requires joint target detection and identity maintenance across complex aerial video sequences. Although global self-attention mechanisms excel at context modeling, their quadratic computational complexity O(N^2) presents severe bottlenecks for deployment on resource-constrained onboard UAV hardware. To bridge this gap, we introduce AeroMamba-YOLO (AM-YOLO), an efficient tracking-by-detection framework that integrates a Mamba-enhanced backbone into an anchor-free YOLO architecture. Specifically, we propose two novel modules, SPPCSMamba and C2CSMamba, which embed linear-complexity O(N) selective state-space modeling into the feature extraction network. These modules adaptively fuse local and global spatial contexts, significantly improving the robustness of feature representations against severe occlusions and scale variations inherent to small targets. Evaluated on a challenging multi-drone RGB–thermal aerial tracking benchmark using BoT-SORT, AM-YOLO delivers state-ofthe-art performance, achieving 69.2 HOTA/95.1 IDF1/90.8 MOTA on Drone 1, and 70.6 HOTA/91.5 IDF1/92.2 MOTA on Drone 2. Comprehensive ablation studies demonstrate the complementary nature of our proposed modules and verify that the performance gains are fundamentally driven by selective state-space modeling rather than a mere increase in parameter capacity
- 4 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessAM-YOLO: Efficient Linear-Complexity Global Context Modeling for UAV Multi-Object Tracking
Multi-object tracking (MOT) from unmanned aerial vehicles (UAVs) requires joint target detection and identity maintenance across complex aerial video sequences. Although global self-attention mechanisms excel at context modeling, their quadratic computational complexity O(N^2) presents severe bottlenecks for deployment on resource-constrained onboard UAV hardware. To bridge this gap, we introduce AeroMamba-YOLO (AM-YOLO), an efficient tracking-by-detection framework that integrates a Mamba-enhanced backbone into an anchor-free YOLO architecture. Specifically, we propose two novel modules, SPPCSMamba and C2CSMamba, which embed linear-complexity O(N) selective state-space modeling into the feature extraction network. These modules adaptively fuse local and global spatial contexts, significantly improving the robustness of feature representations against severe occlusions and scale variations inherent to small targets. Evaluated on a challenging multi-drone RGB–thermal aerial tracking benchmark using BoT-SORT, AM-YOLO delivers state-ofthe-art performance, achieving 69.2 HOTA/95.1 IDF1/90.8 MOTA on Drone 1, and 70.6 HOTA/91.5 IDF1/92.2 MOTA on Drone 2. Comprehensive ablation studies demonstrate the complementary nature of our proposed modules and verify that the performance gains are fundamentally driven by selective state-space modeling rather than a mere increase in parameter capacity
- 11 June 2026
- Published by Springer Nature in International Journal of Speech Technology
- Vol. 29 (2) , 46
A multimodal speech-based pipeline with joint emotion analysis for Vietnamese service quality assessment
- 11 June 2026
- Published by Springer Nature in International Journal of Speech Technology
- Vol. 29 (2) , 46
A multimodal speech-based pipeline with joint emotion analysis for Vietnamese service quality assessment
- Open Access
- 13 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 74030-74047
EfficientEndoSeg: A Lightweight YOLOv5-Based Framework for Endoscopic Surgical Instrument Segmentation and Skill-Related Motion Analysis
This study addresses the challenge of achieving accurate and real-time surgical instrument instance segmentation under strict computational constraints in minimally invasive surgery by proposing EfficientEndoSeg, a compact and high-throughput segmentation framework coupled with tool-tip analysis. Building on the You Only Look Once (YOLO) version 5 instance segmentation baseline, EfficientEndoSeg incorporates four targeted architectural refinements: the C3SG2 module combining GhostNetV2 and Swin Transformer components to enhance local–global representation, Ghost Convolution for parameter-efficient feature extraction, a Double-Attention (A2) block for contextual refinement, and Complete Intersection-over-Union–guided Soft Non-Maximum Suppression to improve robustness in crowded surgical scenes. The framework is systematically evaluated on an internally annotated extension of the m2cai16-tool dataset, comprising 5,249 endoscopic frames and 8,785 polygon-level annotations across seven instrument categories, which serves as a controlled benchmark for experimental validation. These architectural modifications reduce model parameters by 25% and increase inference throughput by 52.7%, while consistently improving recall and mean Average Precision across multiple Intersection-over-Union thresholds. The proposed model matches or surpasses recent YOLO version 7 to version 12 variants and remains competitive with transformer-based and two-stage frameworks, including SegFormer, ISINet, LETNet, Detectron2, and Mask R-CNN. In addition, a training-free tool-tip tracking module based on modified K3M skeletonization enables trajectory, heatmap, and usage analyses, facilitating qualitative skill-related motion assessment without additional learning overhead. Overall, EfficientEndoSeg delivers a reproducible, lightweight, and deployment-ready solution for segmentation-driven surgical scene understanding and instrument motion analysis.
- 13 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 14, 74030-74047
Open AccessEfficientEndoSeg: A Lightweight YOLOv5-Based Framework for Endoscopic Surgical Instrument Segmentation and Skill-Related Motion Analysis
This study addresses the challenge of achieving accurate and real-time surgical instrument instance segmentation under strict computational constraints in minimally invasive surgery by proposing EfficientEndoSeg, a compact and high-throughput segmentation framework coupled with tool-tip analysis. Building on the You Only Look Once (YOLO) version 5 instance segmentation baseline, EfficientEndoSeg incorporates four targeted architectural refinements: the C3SG2 module combining GhostNetV2 and Swin Transformer components to enhance local–global representation, Ghost Convolution for parameter-efficient feature extraction, a Double-Attention (A2) block for contextual refinement, and Complete Intersection-over-Union–guided Soft Non-Maximum Suppression to improve robustness in crowded surgical scenes. The framework is systematically evaluated on an internally annotated extension of the m2cai16-tool dataset, comprising 5,249 endoscopic frames and 8,785 polygon-level annotations across seven instrument categories, which serves as a controlled benchmark for experimental validation. These architectural modifications reduce model parameters by 25% and increase inference throughput by 52.7%, while consistently improving recall and mean Average Precision across multiple Intersection-over-Union thresholds. The proposed model matches or surpasses recent YOLO version 7 to version 12 variants and remains competitive with transformer-based and two-stage frameworks, including SegFormer, ISINet, LETNet, Detectron2, and Mask R-CNN. In addition, a training-free tool-tip tracking module based on modified K3M skeletonization enables trajectory, heatmap, and usage analyses, facilitating qualitative skill-related motion assessment without additional learning overhead. Overall, EfficientEndoSeg delivers a reproducible, lightweight, and deployment-ready solution for segmentation-driven surgical scene understanding and instrument motion analysis.
- 1 May 2026
- Published by Springer Nature
- p. 331-341
A Hierarchical Framework for Real-Time Multi-person Pose Estimation
- 1 May 2026
- Published by Springer Nature
- p. 331-341
A Hierarchical Framework for Real-Time Multi-person Pose Estimation
- Open Access
- 23 April 2026
- Published by European Alliance for Innovation n.o. in EAI Endorsed Transactions on AI and Robotics
- Vol. 5
EmoFedProto: Privacy-Preserving Vietnamese Speech Emotion Recognition via Prototype-Based Federated Learning
Speech Emotion Recognition (SER) plays a fundamental role in affective computing by enabling machines to infer human emotional states from vocal expressions. However, most existing SER systems rely on centralized training paradigms, which raise serious privacy concerns due to the sensitive nature of speech data. Federated Learning (FL) offers a privacy-preserving alternative by allowing collaborative model training without sharing raw data, yet its performance often degrades significantly under non-IID data distributions, a common characteristic of speech emotion datasets caused by speaker variability and emotion imbalance. To address these challenges, we propose EmoFedProto, a prototype-based federated learning framework with clustering-enhanced prototype aggregation tailored for Vietnamese speech emotion recognition in low-resource settings. Instead of exchanging full model parameters, EmoFedProto communicates class-level feature prototypes, enabling more robust alignment across heterogeneous clients. Experiments conducted on the VNEMOS dataset under realistic non-IID and few-shot conditions demonstrate that EmoFedProto achieves an accuracy of 0.875, outperforming the baseline FedProto (0.825), while reducing performance variability by 44%. These results indicate that clustering-based prototype federated learning is an effective and communication-efficient solution for privacy-preserving speech emotion recognition, particularly in low-resource languages and realworld federated environments.
- 23 April 2026
- Published by European Alliance for Innovation n.o. in EAI Endorsed Transactions on AI and Robotics
- Vol. 5
Open AccessEmoFedProto: Privacy-Preserving Vietnamese Speech Emotion Recognition via Prototype-Based Federated Learning
Speech Emotion Recognition (SER) plays a fundamental role in affective computing by enabling machines to infer human emotional states from vocal expressions. However, most existing SER systems rely on centralized training paradigms, which raise serious privacy concerns due to the sensitive nature of speech data. Federated Learning (FL) offers a privacy-preserving alternative by allowing collaborative model training without sharing raw data, yet its performance often degrades significantly under non-IID data distributions, a common characteristic of speech emotion datasets caused by speaker variability and emotion imbalance. To address these challenges, we propose EmoFedProto, a prototype-based federated learning framework with clustering-enhanced prototype aggregation tailored for Vietnamese speech emotion recognition in low-resource settings. Instead of exchanging full model parameters, EmoFedProto communicates class-level feature prototypes, enabling more robust alignment across heterogeneous clients. Experiments conducted on the VNEMOS dataset under realistic non-IID and few-shot conditions demonstrate that EmoFedProto achieves an accuracy of 0.875, outperforming the baseline FedProto (0.825), while reducing performance variability by 44%. These results indicate that clustering-based prototype federated learning is an effective and communication-efficient solution for privacy-preserving speech emotion recognition, particularly in low-resource languages and realworld federated environments.
- 1 April 2026
- Published by Springer Nature
- p. 239-246
MoodFlow: Cross-Modal Music Recommendation for Short-Form Videos via Neural ODEs with Mood-Aligned Music Recommendations
- 1 April 2026
- Published by Springer Nature
- p. 239-246
MoodFlow: Cross-Modal Music Recommendation for Short-Form Videos via Neural ODEs with Mood-Aligned Music Recommendations
- 18 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
Colubrid-Net: A Unified Cross-Modal Framework for Hydrological Forecasting in An Khe Reservoir, Vietnam
Accurate reservoir water level forecasting is critical for water resource management and flood mitigation in monsoon-affected regions like Vietnam. Traditional methods approaches fail to capture complex spatiotemporal dynamics in reservoir systems. This paper presents Colubrid-Net, a novel unified cross-modal architecture that combines satellite imagery and time series data for enhanced water level prediction, to address temporal misalignment between monthly satellite observations and daily measurements, we proposed an interpolation mechanism with seasonal variations. Experiments on the An Khe reservoir data ranging from 2019 to 2022 showing an robust results with MAE of 0.0242, MSE of 0.0022 and RMSE with 0.0464, substantially outperforming traditional baselines, showing the effectiveness of unified multi-modal architectures for hydrological forecasting.
- 18 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 23 (1545598X) , 1-5
Colubrid-Net: A Unified Cross-Modal Framework for Hydrological Forecasting in An Khe Reservoir, Vietnam
Accurate reservoir water level forecasting is critical for water resource management and flood mitigation in monsoon-affected regions like Vietnam. Traditional methods approaches fail to capture complex spatiotemporal dynamics in reservoir systems. This paper presents Colubrid-Net, a novel unified cross-modal architecture that combines satellite imagery and time series data for enhanced water level prediction, to address temporal misalignment between monthly satellite observations and daily measurements, we proposed an interpolation mechanism with seasonal variations. Experiments on the An Khe reservoir data ranging from 2019 to 2022 showing an robust results with MAE of 0.0242, MSE of 0.0022 and RMSE with 0.0464, substantially outperforming traditional baselines, showing the effectiveness of unified multi-modal architectures for hydrological forecasting.