As AI progresses, there is no denying the potential benefits to video security. The AI-powered video analytics market is projected to exceed $133 billion by 2030, from $32 billion in 2025, due to a wide range of applications including urban, retail, logistics, manufacturing, residential development and transportation.
However, to unlock these benefits, organizations must always make mistakes on the part of responsible use. Most importantly, this has also been extended to the data used to power and train AI models.
Understanding data transparency and quality
As highlighted by the recent Amnesty International Report, the use of AI-powered video surveillance needs to be transparent.
However, to do so, it is important that the AI model return to the stage and carefully considers the quality and origin of the data that has been trained and fine-tuned. With AI-enabled videos being deployed, developers cannot fall into the same trap as their colleagues working on the large-scale language models (LLM) on which the generator AI relies. In other words, the challenge is to procure sufficient data for AI model training. Unfortunately, LLM developers find themselves paying an increasing amount for a large dataset, due to the lack of privacy measures or litigation from the rights holder. Bias from non-representative data could have had knock-on impacts on the trust of AI people and polluted the implementation of AI, which provides insights.
Building an accurate video AI model
To alleviate many of these challenges before spreading it, it is important to consider how to obtain access to high quality, responsible procured visual data to train AI video models. The datasets used to train AI models are representative and diverse to ensure accuracy and fairness, and must be legally sourced to respect the IP rights of data owners. This is not an easy task to retrieve, especially when dealing with sensors such as cameras that can collect a lot of personal and sensitive information.
One solution to this challenge is Project Hafnia. Exploiting the Nvidia Nemo Curator and AI Models, a platform developed by Milestone Systems in collaboration with Nvidia. With Project Hafnia, data generators share and utilize data, allowing developers to access traceable, regulatory annotated video data they can use to train AI models.
One of the platform's first data generators is an American city in Dubuque, Iowa. In addition to AI Analytics Company Vaidio, Milestone has built a joint visual language model that improves AI accuracy from 80% to over 95%, via anonymization and curation of Dubuque's raw videos. This leap has enhanced smarter traffic management, quicker emergency response, and public safety. Everything is done responsibly and without a massive infrastructure overhaul.
Milestone, a company specializing in anonymization solutions, recently acquired Bright AI, has added an additional layer of data privacy to project Hafnia. Therefore, bright AI technology automatically detects personal identifiers such as faces and license plates and generates synthetic exchanges.
Integrating and curating data from multiple data generators is one way for developers to obtain sufficient visual data to develop accurate AI models for detecting events such as vandalism, vehicle accidents, and traffic flows.
Composite data for difficult-to-assemble data sets
Another solution is provided in the form of synthetic data, which is an artificially generated or extended data set that simulates or generalizes real conditions. Using synthetic data, AI developers can train models with a vast amount of diverse and representative information, while reducing ethical and legal concerns surrounding privacy and consent.
For example, in Aalborg Harbor in Denmark, it was impossible to train AI models to detect individuals falling into ports due to the dangers posed to human volunteers. The dataset also had to include a variety of human actors, such as wheelchair users. Using dummies also did not fully capture the complete complexity. Therefore, the best solution was synthetic data that avoided safety and ethical concerns and could extend the training dataset in a variety of fall scenarios. The AI model developed through this process shows promising results warning rescue teams if a person falls into port, minimizing response times and increasing the chances of survival by reducing cold water exposure.
Unlock the possibilities of AI with video
AI has high expectations for the safety of our cities, buildings and individuals. However, this can only be achieved with AI models that fully capture the complexity of the built environment and human behavior. Video analytics developers should consider options when trying to build a comprehensive dataset for AI model training. New responsible options are emerging – from consolidated data collected by many data generators to synthetic data generators. It's the question of where to look.
