Abstract

Visual impairment and age-related decline significantly affect independent mobility and safety in daily life. This paper presents a smart monitoring and assistance system aimed at supporting visually impaired individuals and elderly people through the integrating of artificial intelligence, computer vision, and sensor-based technologies. Getting inspired by the working mechanism of the human eye, the proposed system employs the YOLO11 deep learning model for real-time object detection and classification, combined with the Depth Anything v2 model for monocular depth estimation to calculate the distance between users and surrounding objects.

The system is implemented using an embedded camera and IoT-based sensors, including ultrasonic distance sensors, GNSS positioning, heart rate monitoring (MAX30102), and fall detection modules, enabling comprehensive environmental perception and user health monitoring. Experimental evaluations were conducted in both bright and low-light environments using a self-collected dataset. The results depict that the proposed system achieves an overall the accurate of detection -- approximately 95 percent, with stable performance across varying lighting conditions.

The findings confirm the feasibility and effectiveness of integrating deep learning models with embedded hardware to provide real-time assistance. This system has strong potential for development into a wearable smart device capable of enhancing mobility, reducing collision risks, and improving the independence and quality of life for visually impaired and elderly users. Moreover, the proposed approach contributes to the advancement of human-centered intelligent assistive technologies with meaningful social impact.

Abstract

Ho Chi Minh City AI Challenge 2025, now in its sixth edition, focused on advancing event retrieval techniques from large video collections, driving research and innovation in video analysis. The challenge featured a dataset of 1,471 videos spanning 328 h alongside diverse query formats to evaluate system performance in realistic scenarios. Participant teams competed in multiple rounds, addressing complex queries involving temporal and semantic event understanding. Leveraging advanced deep learning models, temporal segmentation, and multimodal fusion techniques, participants showcased innovative approaches across textual and visual Known-Item Search and Question Answering tasks. Visual KIS recorded the highest performance, highlighting the advantages of rich visual context over text-based queries. This paper provides an overview of the challenge organization, dataset, methodologies, evaluation metrics, and insights into trends and solutions observed during the competition.

Abstract

This paper presents an end-to-end AI-based bill and receipt recognition system that integrates object detection, optical character recognition, and an interactive graphical user interface for efficient analysis and comparison. The proposed pipeline employs YOLOv8 to accurately detect and localize bill regions from input images, followed by a high-accuracy OCR model as the primary recognition engine. PaddleOCR is incorporated as a baseline to enable cross-validation and performance comparison between OCR approaches.

To enhance usability and transparency, the system is deployed through a modern PySide6-based graphical interface featuring real-time visualization of detection results, detailed tabular OCR outputs, and confidence-based analytics. The GUI supports side-by-side comparison of OCR models, color-coded confidence indicators, and structured export of recognition results in JSON format, including bounding boxes and metadata. Experimental evaluation demonstrates that the primary OCR model achieves higher recognition confidence with fewer but more reliable text predictions, while PaddleOCR provides broader text coverage with slightly lower confidence. The proposed system effectively balances accuracy, usability, and extensibility, making it suitable for practical bill digitization, analysis, and future batch-processing applications.

Abstract

This study presents the development of an intelligent STEM-oriented system designed to support school health initiatives through the identification and analysis of natural compound extracts. The system integrates Artificial Intelligence (AI) and computer vision technologies to automatically recognize solid and liquid substances derived from natural materials such as ginger, citrus fruits, and plant-based oils. Using an object detection model deployed via ONNX Runtime Web, the system processes user-uploaded images and classifies extracted compounds in real time.

The proposed solution aims to enhance STEM education by connecting scientific experimentation with practical health-related applications. By enabling students to identify bioactive substances from natural sources, the system promotes awareness of sustainable materials and their benefits in immunity support, antimicrobial activity, and overall wellness. In addition, the platform provides descriptive insights into detected compounds, including their chemical characteristics, biological functions, and potential applications in food, cosmetics, and traditional medicine.

Experimental implementation demonstrates that the system operates efficiently with high recognition reliability and fast processing time, making it suitable for classroom environments and interactive learning scenarios. Beyond education, the model shows potential for broader applications in smart laboratories and green technology initiatives.

This research contributes to the integration of AI-driven visual recognition into STEM-based health education, fostering innovation, environmental awareness, and interdisciplinary learning among students.

Abstract

During peak hours, congestion frequently occurs at school parking entrances due to the large number of students arriving simultaneously to deposit their vehicles. Manual ticket distribution slows down the process, leading to crowding, disorder, and potential safety risks. To address this issue, this project proposes an automatic vehicle ticket dispensing machine designed to improve the efficiency of school parking management.

The system operates by detecting the presence of a user through object detection sensors and automatically releasing a parking ticket using a servo-controlled mechanism. A microcontroller-based control system processes sensor input and executes the ticket dispensing action without human intervention. The model is designed to be simple, cost-effective, and practical for real-world school environments.

By reducing waiting time, minimizing human error, and streamlining the ticketing process, the proposed system contributes to alleviating traffic congestion at school gates. This project also demonstrates the application of engineering and technological knowledge in solving practical problems within educational settings, promoting a more modern, efficient, and convenient parking management approach.