BPA Tech on the development of computer vision and VLMs for GlobalMedia

Co-founder and Chief Commercial Officer of BPA Tech Arkady Froymchuk served as an expert in a new feature by the GlobalMedia portal network, focusing on areas that saw significant growth in demand during the third quarter of 2026.

In the article, experts from various industries shared which products and technologies gained momentum from July to September, what drove this growth, and which areas companies plan to scale further.

BPA Tech operates in the IT sector, primarily focusing on artificial intelligence and computer vision systems, so we have a clear view of how business demand for such solutions is changing. Over the past year, the market has noticeably shifted from an interest in isolated AI features to more comprehensive systems capable of processing large volumes of data and solving less formalized tasks.

From July to September, the company saw a notable increase in demand for applied, comprehensive AI systems, particularly in computer vision and intelligent video analytics. We develop these solutions for business, industry, and urban infrastructure, so it is especially telling for us how the nature of client requests is evolving. While most tasks used to be limited to the automatic monitoring of predefined events, companies now increasingly want systems that can analyze video streams with much greater flexibility and respond to non-standard user queries.

One area that has been developing particularly actively in recent months is VLM (Visual Language Models). These are multimodal neural network models capable of working with visual information and natural language simultaneously. In effect, they allow a person to assign a task to the system not through a pre-programmed algorithm or a fixed set of classes, but through a simple text prompt.

This fundamentally expands the capabilities of classic video analytics. Traditional computer vision works well where it is clear in advance exactly what needs to be monitored: the presence of a hard hat or vest, a vehicle entering a specific zone, the appearance of smoke, or the number of people in a room. For each such scenario, an event is defined in advance for the system to detect. VLMs allow for a different category of tasks. For example, when it is impossible to describe all potential situations in advance, but it is necessary to quickly find a specific object, person, or event within a large volume of video data.

We see particularly great potential here in security. For example, a child might go missing in a public space—a park, a shopping mall, a transit hub, or at a large event. In a classic scenario, security personnel are given a description: age, clothing, jacket or shirt color, and whether they have a backpack. The operator then has to manually switch between cameras and review footage, trying to find a matching person. As we know, this can take a significant amount of time, and in some situations, every second counts.

With the use of VLMs, this process can look different. The operator gives the system a prompt, such as: "find a boy about eight years old in a yellow t-shirt with a red backpack." The model analyzes images from available cameras and highlights fragments that match the description. Once the person is detected, the system can help determine which camera they appeared on first, where they moved next, and in which part of the area they were last seen, or show exactly where they are right now. In other words, this is no longer just about detecting a person in a frame, but about searching via semantic description and analyzing a sequence of events.

A similar mechanism is applicable to other security tasks. For example, you can search for an unattended item by its description, find a person in specific clothing, analyze a person's appearance in multiple zones of a facility, or quickly find a necessary episode in hours of video archives. Of course, this applies not only to people but also to animals and any other objects. This is especially important for facilities with a large number of cameras, where manual review of footage becomes too time-consuming and labor-intensive.

Another typical scenario is searching for events in archived footage. For example, a visitor to a restaurant or shopping center reports losing a bag but does not remember the exact time or place. Instead of sequentially watching several hours of video, an employee can formulate a query: find a person with a specific description and the moment when a bag appears or is left next to them. The system narrows the data set down to a few of the most relevant episodes, which the operator can then verify.

This is precisely why we view VLMs as the next stage in the evolution of computer vision systems. The main change is that a video archive ceases to be just a collection of recordings and becomes data that can be searched using natural language.

In classic computer vision, the scenario is usually set in advance. We define beforehand exactly what the system should look for: a person, a car, a hard hat, smoke, a suitcase, crossing a zone, and so on. The model, rules, and event logic are configured for this task. For example, for an airport, you can set a scenario in advance: if a suitcase appears in a certain zone and no person is near it for five minutes, the system triggers an alarm. This approach works well when the event is clearly formalized and it is known in advance what needs to be monitored.

VLMs differ in that they allow for working with a more flexible description of a situation. That is, it is not necessary to pre-program a separate scenario for every possible case. The operator can formulate a query after the event has occurred: for example, "find the moment when a man in a dark jacket left a blue suitcase near the check-in counter and walked away." The model analyzes not only the presence of individual objects but also their relationship, actions, and the context of the scene.

This is why VLMs are particularly interesting for security tasks. For example, at an airport, you can set up basic rules in advance: unattended luggage, unauthorized entry into a restricted zone, or overcrowding. But the real situation is almost always more complex than a pre-written template. You might need to find a specific episode based on an eyewitness description, understand who brought an item, where the person was before that, who they interacted with, and on which camera they appeared later. For classic CV, each such scenario would have to be formalized separately. VLMs allow you to ask such questions in natural language and search for the corresponding fragments in the video archive.

We see the greatest potential in a hybrid architecture: classic computer vision continuously monitors predefined scenarios, while VLMs are engaged where it is necessary to understand context, perform non-standard searches, or quickly analyze an event that could not have been formalized in advance.

Therefore, we associate the further scaling of this area primarily with urban and public safety, industrial facilities, retail, and systems for managing large video archives. For us, this is one of the areas where, over the past few months, not only has technological interest grown, but so has the number of specific applied scenarios from our clients.

‍