Young Vietnamese engineers top global AI playground for two consecutive years

Recently, the AI City Challenge 2025 (AI in smart cities) results announcement ceremony took place within the framework of the International Conference on Computer Vision (ICCV 2025) in Hawaii (USA). Following the victory in 2024, this year VNPT 's AI engineering team won first place in the category of Processing and recognizing objects from ultra-wide-angle camera image data at edge devices (edge AI). The problem requires an AI system with real-time processing speed, directly on compact hardware devices, while ensuring high accuracy in recognizing objects from highly distorted image data, meeting practical needs.

AI City Challenge 2025 is one of the world's most prestigious annual competitions on artificial intelligence (AI) applied in smart cities. This year's competition includes four categories with higher complexity than previous seasons, attracting more than 30,000 teams from countries with strong AI development such as the US, China, Korea, Taiwan, etc.

VNPT's engineering team leads the way thanks to AI's accuracy and processing speed at the edge.

The challenge of enhancing AI capabilities at the edge

The problem of processing and recognizing objects from ultra-wide-angle camera image data has been included in the AI City Challenge since 2024, reflecting the trend of applying computer vision to current traffic monitoring systems. With high practicality, the category is always the race with the largest number of teams in the whole competition. This year, the difficulty of the category is increased when, in addition to accurately processing distorted and deformed images, teams must optimize the entire process so that the model can operate effectively on edge devices.

VNPT engineering team applies many techniques to quickly and accurately identify vehicles.

Teams were required to optimize their models to run efficiently on the Jetson Orin, a small device located at the data collection point (called an edge device), which has a power limit of 30W and much lower computing power than a central server. This meant that teams could not use overly large models, but had to streamline and optimize the program to run faster, consume less resources, and still recognize vehicles accurately. These changes made the AI City Challenge 2025 one of the most difficult seasons to date, especially as teams learned from last year's experience and the level of competition increased significantly.

Benefit from real-world model optimization experience

In traffic monitoring problems, computing infrastructure and network connections are often limited, making the development of AI models that are both accurate and efficient a big challenge. This is also the reason why Edge AI has become an inevitable trend. Instead of sending all data to a central server for processing, the model is placed right at the collection device (such as a camera), helping to respond faster, reduce latency, save bandwidth and ensure data security, especially in large-scale monitoring systems.

According to the results announced at the AI City Challenge 2025, the VNPT engineering team came in first, surpassing hundreds of teams from major technology corporations, research institutes and universities around the world. This achievement contributes to strengthening the AI ecosystem for traffic monitoring and urban security in the country, where ultra-wide-angle cameras are being widely deployed to expand the observation area, reduce blind spots, reduce the number of installed devices and improve the efficiency of urban infrastructure operations. With more than seven years of experience developing AI models for image processing and deploying them domestically, the VNPT engineering team has accumulated the ability to balance accuracy, speed and operating costs - factors that determine the effectiveness of applying AI in real environments.

VNPT team has developed and mastered more than 40 image processing AI models.

Currently, the VNPT team has developed and mastered more than 40 different image processing AI models such as license plate recognition, traffic flow measurement, helmet detection, as well as models specific to Vietnam such as detecting vehicles carrying three people, carrying bulky goods, or detecting fires and weapons in the field of security and urban surveillance. These models are optimized to operate on a variety of hardware, from GPUs, CPUs to NPUs, meeting the diverse requirements of systems and customers.

To be able to deploy effectively on a large scale, especially in the on-premise model and at the edge with hundreds of cameras simultaneously, VNPT engineers have also built optimal processing methods that allow the simultaneous operation of hundreds of video data streams. This approach makes AI solutions easy to scale, saves resources and is suitable for infrastructure conditions in many localities.

Applying that experience to the AI City Challenge 2025, the team applied a combination of techniques such as model compression to reduce size and processing resources, image processing flow optimization to reduce latency, and programming language and code structure refinement to form the overall processing chain for the highest performance. This approach helps the model maintain accuracy while increasing inference speed and deployment on hardware-constrained edge devices.

Multi-disciplinary deep learning AI research platform

Not only possessing a team of young, talented staff and strong computing infrastructure, one of the important premises for VNPT to have a comprehensive AI product ecosystem that meets international standards and wins high prizes at prestigious competitions is the in-depth AI research platform in many fields.

Regarding AI applications in image processing, in addition to smart urban and traffic systems, VNPT also promotes research applications in the medical field. In September 2025, the group announced scientific research at MICCAI 2025 - the world's leading conference on AI and computer vision in medicine. The research focused on AI applications in thyroid cancer diagnosis, conducted with data from nearly 10,000 patients in 3 regions of the country over 4 years. The project plays a fundamental role in developing automatic diagnostic support systems suitable for the population characteristics and medical conditions in the country, helping to improve accuracy, shorten diagnosis time, reduce workload for doctors, and expand access to high-quality medical services to grassroots levels.

VNPT's AI engineering team published research at many prestigious scientific conferences.

In the field of language and speech processing, VNPT also announced research at EMNLP 2025 - an A*-ranked AI conference on natural language processing, and ICASSP 2025 - an A1-ranked conference on speech processing. These works apply advanced machine learning techniques in developing large language models (LLM) and generative AI, aiming to improve the ability of AI models to understand Vietnamese context, emotions and nuances - which are complex factors and rarely modeled in international corpora. Thanks to that, the Vietnamese language is increasingly present in the global NLP research community, opening up the possibility of building strong, unique and suitable language models for Vietnamese people.

Source: https://vnpt.com.vn/gioi-thieu/tin-tuc/nhom-ky-su-tre-viet-nam-hai-nam-lien-tiep-dung-dau-san-choi-ai-toan-cau.html