NEC Builds VLM-Based AI That Scores Camera Footage to Spot Disaster Signs

Scoring conditions instead of reading text answers

NEC says a general-purpose VLM can describe on-site conditions in words, but its text responses do not carry quantitative information such as degree or extent. That makes it hard to capture subtle differences between scenes. Asking a VLM a text-based question, for example whether an area is flooded, can also produce answers that diverge from the facts or are inconsistent, according to the company.

NEC's approach does not rely on the text the model generates. Instead, it uses proprietary technology to analyze the likelihood values the VLM computes internally while generating text. NEC says this draws out information that would otherwise be lost when it is converted into words, and allows fine scoring of conditions such as the extent of flooding or snow accumulation. Changes and trends over time can then be tracked as continuous numerical values. Operators can set precursor and danger thresholds suited to each site and disaster type based on these scores.

NEC explains the underlying method as recognizing and quantifying subtle differences between moving images by comparing the sentence-generation likelihoods the VLM calculates when describing them. The work is scheduled to be presented at the international conference ACM Multimedia Asia 2026 (MMAsia 2026).

Prompts generated automatically for each site

A second problem NEC addresses is prompt design. Because a VLM's answers are heavily influenced by the content of the prompt, reflecting the points to watch at each site normally requires engineers with specialized knowledge to work through trial and error. NEC says this has made it impractical to apply the approach quickly across many disaster types and locations.

With the new technology, AI automatically generates a prompt suited to the disaster type being observed, such as flooding or snow accumulation, and to the composition of the scene captured by each camera. The prompt is built from footage from the site through successive adjustments that keep the score continuous between temporally preceding and following images. In the process, the system automatically identifies which areas of the image the VLM should focus on and which can be disregarded.

According to NEC, this suppresses score fluctuations caused by minor differences between images and enables stable judgments. Experts no longer need to design or adjust prompts for each camera's installation environment or disaster type.

Background and what it means for monitoring sites

NEC points out that many roads and rivers already have live cameras, but continuously checking footage from numerous cameras by hand is difficult, and subtle changes that signal an approaching disaster can be overlooked. Image recognition AI is one option for automation, but the company notes it requires collecting and annotating large volumes of training data for each disaster type and site, which is time-consuming and labor-intensive.

The new technology can use existing live cameras operated by local governments, government agencies and infrastructure operators, which NEC says makes it easy to deploy at monitoring sites. The company expects it to streamline monitoring operations and speed up on-site confirmation and response decisions, helping prevent damage and limit its spread. NEC says it will use the technology to help local governments and infrastructure operators and to contribute to greater safety and security in local communities.