JMST's Section on Computer Science and Control Engineering.

A smooth robust controller for the path following of underactuated surface vehicles under disturbances

Duc Sang Cao — 2025-12-31

Trajectory tracking for underactuated surface vehicles (USVs) is challenging due to disturbances, system uncertainties, and underactuation. This paper proposes a smooth robust controller for USVs operating under such conditions. A translational velocity-based guidance law is designed, combined with a control handpoint technique [9] and the introduction of a small auxiliary term into the tracking error. A smooth adaptive force controller is developed to drive the USV to follow the desired trajectory. Unlike sliding mode control, the proposed controller produces continuous control signals and does not require prior knowledge of disturbance bounds. Lyapunov stability theory and Barbalat’s lemma are employed to analyze stability. Simulations in Matlab–Simulink for different values of ε, compared with the controller in [17], demonstrate that the USV tracks the desired trajectory with a small error while maintaining smooth control forces and moments.

Research integration of a real-time object detection model on an underwater observation system using Laser Range-Gated imaging

Le Van Hoang — 2025-12-31

In recent years, underwater observation technology utilizing Laser Range-Gated (LRG) imaging has garnered considerable attention. This is attributed to its capability to provide high-resolution imagery in low-light environments, particularly in deep-sea settings at depths of hundreds of meters, which are characterized by the absence of light and low water transparency. However, practical deployment remains challenging due to significant light attenuation and scattering within the water medium. This paper presents a study on the integration of a real-time object detection model into an LRG underwater observation system. This integration is based on the analysis of reflected laser signals and adaptive image processing. The system utilizes a pulsed laser illumination source operating at a 532 nm wavelength. Image acquisition is performed by a high-sensitivity gated Intensified CCD (ICCD) camera, which is synchronized with the emitted laser pulses. The processing framework is built upon Python–OpenCV. Experimental results demonstrate that the system operates stably, clearly detecting the reflective regions of the target image. It automatically adjusts thresholding based on background illumination and achieves real-time performance at a rate of 28 - 30 frames per second (FPS).

An enhanced fuzzy time series forecasting model using Gaussian membership functions and PSO-based parameter optimization

Tinh Nghiem Van — 2025-12-31

In the context of data-driven decision making and increasing market volatility, accurate time series forecasting plays a crucial role in sectors such as education and energy. This paper presents a novel fuzzy time series (FTS) model that integrates Gaussian membership functions with Particle Swarm Optimization (PSO) to simultaneously optimize the universe of discourse ( , ) and the standard deviations of Gaussian functions. The key innovation lies in the combination of dynamic fuzzification with PSO-based parameter optimization, effectively addressing the limitations of static partitioning and manual tuning in traditional FTS models. Furthermore, the model employs time-dependent fuzzy logical relationship groups (TD-FLRGs) to improve forecasting accuracy in handling non-linear and uncertain data. The proposed model is evaluated on the Alabama university enrollment dataset (1971–1992) and validated using Vietnam's RON95 gasoline prices. With seven intervals and first-order relationships, it achieves a Root Mean Squared Error (RMSE) of 318.7. These results highlight the model's superior performance and demonstrate its potential as a scalable and adaptable solution for real-time forecasting in dynamic environments.

Integrated INS/GPS navigation solution for high-speed vehicles

Dr Khoi Nguyen Van — 2025-12-31

Experimental results indicate that the noise characteristics of micromechanical sensors vary significantly with environmental and operational conditions, reducing the effectiveness of conventional Kalman filtering. For high-speed aerial vehicles, GPS not only provides accurate positioning information but also a highly precise heading measurement (Heading <0.3°). This paper proposes an integrated Inertial Navigation System (INS) and Global Positioning System (GPS) solution employing a Strongly Robust Adaptive Kalman Filter (SH-RAKF) to enhance positioning and orientation accuracy. State and observation models are constructed based on experimental data obtained from micromechanical sensors and GPS. Simulation results demonstrate that the proposed algorithm significantly improves trajectory tracking and heading estimation under conditions of strongly time-varying noise.

Research on solutions to accelerate computation for CNN executed on resource-limited SoC-FPGA

Hai Nguyen — 2025-12-31

This paper presents a method for designing and implementing an image recognition model on a System on Chip (SoC) platform integrated with Field Programmable Gate Arrays (FPGA). With the increasing demands for low-latency inference processing in edge computing and high energy efficiency, customized hardware acceleration solutions become essential. The embedded execution results of the model on the FPGA hardware of the SoC are compared with the PyTorch model running on the same resource-limited FPGA hardware, showing that this architecture provides a high-performance, low-latency solution, and demonstrates the feasibility of using Vitis High-Level Synthesis (HLS) to quickly generate specialized IP cores for edge computing applications. The results will serve as a solid foundation for continuing to build more complex network models for larger problems.

Two-stage fine-tuning of whisper with metric-driven dataset selection for domain-specific Vietnamese ASR

Duc Thinh Dang — 2025-12-31

Automatic speech recognition (ASR) in low-resource, domain-specific settings remains challenging due to limited labeled data and domain mismatch. This paper proposes a framework that combines metric-based donor dataset selection with a two-stage fine-tuning strategy to adapt the Whisper model for Vietnamese military-specific ASR. The Fréchet DeepSpeech Distance (FDSD) metric is used to identify the most acoustically and phonetically similar general-domain dataset to the target Military Information Retrieval (MIR) corpus. VN-SLU dataset was selected for Stage 1 fine-tuning, bridging the domain gap before Stage 2 fine-tuning on MIR for domain specialization. Experimental evaluation on the MIR test set shows that the proposed method achieves a Word Error Rate of 3.49% and a Character Error Rate of 2.41%, outperforming direct fine-tuning and blended-data approaches. Loss curve analysis confirms that Stage 1 adaptation accelerates convergence and mitigates overfitting in Stage 2. These results demonstrate that integrating metric-driven general-domain dataset selection with sequential fine-tuning is an effective and reproducible approach for enhancing ASR performance in low-resource, domain-specific scenarios.

A simulation-to-real image translation approach via diffusion model for military vehicle recognition

Dr Hong Le Thi Thu — 2025-12-31

The research and development of automated systems for military vehicle recognition are crucial for enhancing situational awareness of commanders, thereby improving combat effectiveness and mission accomplishment. To apply computer vision technologies in such systems, large-scale and diverse training datasets are required, encompassing images of objects captured across varying temporal and spatial conditions. However, in practice, military vehicle imagery, particularly of adversary assets, is difficult to acquire, limited in quantity, and costly in terms of time and resources. Synthetic simulation-based data provide a cost-effective alternative, but the visual gap between simulated and real-world images hinders model performance in real deployments. In this paper, we propose a diffusion-based data generation method for simulation-to-real domain adaptation, enabling the synthesis of realistic, labeled images from simulation data for training recognition models. Specifically, we develop an image generation framework using Flex.2 diffusion model quantized with fp4, guided by edge maps extracted through the Canny filter. Experimental results on a military vehicle dataset demonstrate a substantial improvement in image quality, with the Fréchet Inception Distance (FID) reduced from 270 to 162 compared to real-world imagery. These findings highlight the scalability and flexibility of our approach as a practical solution for simulation-to-real image translation, ultimately improving the generalization and reliability of military vehicle recognition models.

A timing efficient method for interactive multi-objective optimization algorithms

Thu To — 2025-12-31

Multi-objective optimization has been increasingly applied in many real-world domains. Evolutionary algorithms are widely adopted because they can approximate diverse sets of trade-off solutions in a single run. However, the output of these algorithms is a set of Pareto-optimal solutions, and selecting the most appropriate solution depends heavily on the decision maker (DM). In interactive multi-objective evolutionary optimization, the DM’s feedback can also influence the search trajectory, guiding the population toward preferred regions without imposing hard constraints. Nevertheless, incorporating user preferences during evolution must be performed carefully. Multi-objective evolutionary algorithms rely on a delicate balance between convergence and diversity, as well as between exploration and exploitation. Poorly timed interactions may disrupt this balance, causing loss of diversity or premature convergence. This paper proposes a timing-suggestion mechanism that identifies when user interaction is most effective. The method analyzes population dynamics and quality indicators to determine appropriate moments for incorporating DM feedback, ensuring that interaction enhances, rather than destabilizes, the search process. Experimental results on interactive multi-objective optimization algorithms demonstrate that the proposed approach improves interaction effectiveness while maintaining competitive optimization performance.

Deep multi-view fuzzy consensus with uncertainty

Pham Van Nha — 2025-12-31

Clustering multi-view data is challenging due to feature heterogeneity, inter-view inconsistency, and inherent uncertainty. Traditional fuzzy clustering methods (FCM, PFCM) cannot exploit complementary information, while most multi-view models overlook uncertainty and adaptive weighting. We propose a unified deep fuzzy framework named DMFCU (Deep Multi-view Fuzzy Consensus with Uncertainty), which integrates multi-view autoencoder reconstruction, fuzzy clustering in a consensus space, cross-view alignment, uncertainty regularization, and entropy-based view weighting. The optimization alternates updates of memberships, centroids, and view weights with backpropagation for representation learning. Experiments on benchmark datasets show that DMFCU consistently outperforms state-of-the-art fuzzy clustering approaches in accuracy, NMI, and robustness under noisy or incomplete views. The framework achieves strong modeling capacity with comparable complexity, offering a principled solution for reliable multi-view clustering under uncertainty.

A simulated-crowd guided augmentation training framework for realistic crowd-aware robot navigation

Minh Dang — 2025-12-31

Navigating autonomous robots through dense human crowds remains a complex challenge due to the dynamic and socially aware nature of pedestrian behavior. While deep reinforcement learning (DRL) has enabled promising advances in crowd-aware navigation, most existing methods are trained exclusively on synthetic simulations, limiting their generalization to real-world environments. In this work, we propose Synthetic Crowd Generation with Augmentation (SCGA), an augmented training framework that bridges the gap between simulated learning and real-world pedestrian dynamics. SCGA incorporates motion features extracted from a real-world trajectory dataset into the simulation-based DRL training loop. By dynamically enriching synthetic environments with realistic motion patterns, SCGA allows the navigation policy to learn from real human behavior without the need for costly and risky real-world training. We validate our approach through extensive experiments on crowd navigation tasks. Results show that models trained with SCGA exhibit improved safety and performance in both simulated and real-world-inspired settings. Our framework not only enhances the training process but also offers a reusable environment for developing and benchmarking future navigation models. This work presents a practical and effective pathway for improving the real-world applicability of DRL-based navigation systems.

BrainFL: Federated learning for brain diseases classification

Tuan Dung Kieu — 2025-12-31

In the healthcare domain, data privacy is a critical concern. As a result, a recent strategy for training AI models for healthcare-related problems is federated learning (FL), where models are trained locally at the clients, and only their encoded weights are sent to a central server for aggregation. However, data collected from various clinics or organizations may differ significantly in terms of quality, quantity, and distribution. In addition, the availability of clients participating in the training process is often inconsistent. Furthermore, the characteristics of specific diseases may strongly influence the performance of FL algorithms. In this paper, we focus on the classification of brain diseases using medical images. We design a framework, namely called BrainFL, to investigate several FL algorithms (FedAvg, FedNH, and FedProto), coupled with lightweight CNNs (two custom-designed networks named BrainCNN-2 and BrainCNN-4, and the standard ResNet-18) for feature extraction and classification. Our objective is to evaluate three key factors that significantly impact the overall performance of the task: local data distribution, client availability, and disease variation. Experiments are conducted on two benchmarks of brain disease images: ICH and Brain Tumors, under various experimental settings (e.g., client participation rates and levels of non-IID data). The results reveal valuable insights for deploying such FL frameworks in practical applications.

Ship detection and tracking based on an improved YOLO11n and ByteTrack framework

Bac Nguyen Xuan — 2025-12-31

Ship detection and tracking are critical components of intelligent maritime and coastal surveillance systems. This study introduces a robust method for multi-class vessel detection and real-time tracking. The approach improves the YOLO11n model by incorporating the Separated and Enhancement Attention Module (SEAM) and the Cross-scale Channel Fusion Module (CCFM) into the neck, which enhances multi-scale feature aggregation and edge attention. Experimental results demonstrate an increase in mAP50 from 94.26% to 94.84% and in mAP50–95 from 69.59% to 70.40% on a custom ship dataset. The model size increases only slightly, while real-time inference speed remains at 25 FPS on the Jetson Xavier NX. A custom dataset comprising over 7,600 manually labeled images was developed, covering six vessel categories: cargo ship, passenger ship, military ship, sailboat, fishing boat, and patrol ship. Images were sourced from various public repositories to ensure diversity in vessel size, viewing angle, background, and lighting conditions. The complete system integrates the enhanced YOLO11n with the ByteTrack algorithm for real-time vessel detection and tracking. Experimental results confirm the system's feasibility for practical maritime surveillance applications.

POSW-Vote: A precision-oriented weighted voting framework for robust information extraction from domain-specific reports

Hoang Van Toan — 2025-12-31

Information extraction (IE) from unstructured or semi-structured reports remains a challenging task in specialized domains such as military situation reporting, where textual content is narrative, irregular, and context-dependent. Traditional rule-based or named-entity-recognition (NER) methods often fail to achieve sufficient coverage or adaptability in such settings. In comparison, large language models (LLMs) have shown strong potential for schema-based extraction, their outputs exhibit variability across runs and models, limiting consistency and precision. This paper proposes POSW-Vote (Precision-Oriented Similarity-Weighted Voting) — a semantic voting algorithm designed to consolidate multiple LLM outputs into a single, stable, structured representation. The method jointly employs similarity-based clustering, reliability weighting, and superstring-aware selection to identify the most complete and contextually correct information for each schema-defined field. Extensive experiments on real-world, expert-annotated Vietnamese military reports demonstrate that POSW-Vote consistently improves Precision and F1-score compared to single-run and intra-model baselines, while maintaining robustness across heterogeneous models. The results highlight that the proposed framework enhances the stability and reliability of LLM-based extraction without retraining, offering a scalable, model-agnostic solution for high-stakes domains such as defense intelligence and situational monitoring.

Research and proposal of a hybrid quantum artificial intelligence model for classifying marine mammal signals based on passive sonar principles

Dr Hoang Bach Nhat — 2025-12-31

The convergence between artificial intelligence (AI) and quantum computing is ushering in a new era for problems requiring high levels of computational complexity, particularly in digital signal processing. This paper presents an overview and evaluation of the development trends in hybrid Quantum-AI models, focusing on applications for the recognition and classification of underwater acoustic signals—a challenging domain due to instability, high noise levels, and data scarcity. The study proposes a Hybrid Quantum-CNN model, utilizing CNN for feature extraction and dimensionality reduction, combined with a Variational Quantum Classifier for classification in Hilbert space. The results of the proposed model are compared with an equivalent CNN network, demonstrating that HQC achieves higher accuracy (an increase of 3,5%), while excelling in efficiency: faster convergence speed, and superior generalization capabilities on the same real underwater acoustic dataset. These findings highlight the benefits of quantum approaches not only in accuracy but also in constructing more efficient and robust models; hybrid models open promising avenues for applying quantum machine learning to complex signal reception systems in practical.

Vietnamese speech command recognition on microcontroller for UGV control

Le Anh Quang — 2025-12-31

Deep learning on edge computing devices is a feasible approach that not only meets computational efficiency and latency requirements but also provides advantages in terms of security, bandwidth efficiency, and scalability. This paper presents the deployment and execution process of a Vietnamese short command speech recognition task on an ARM Cortex-M microcontroller, targeting applications in autonomous unmanned ground vehicle (UGV) control. Experimental results show that an int8-quantized CNN model for short command recognition achieves an accuracy of 94.7% on ARM Cortex-M4 hardware, with an execution time of only 15 ms. These results demonstrate the feasibility of real-time Vietnamese speech command recognition for UGV control. Furthermore, the findings open up promising directions for deploying deep learning models on ultra-resource-constrained edge devices (EDGE-AI) for practical real-world applications in the future.