Artificial intelligence technologies have long been used in face recognition system, and search for people by using surveillance cameras. However, this is not the only parameters by which to search. For example, a group of researchers from India have trained an artificial intelligence to search for the right people based on their height, gender and the clothes they wear.
This technology may seem to someone very strange, because “getting to know” people can get more accurate data. But this is not true. The researchers cite an example. Imagine that you only know certain search parameters and the approximate location. And instead of study all the material from all cameras, you can create a query on, for example, “women in red shirts, whose height 153 cm”. It will narrow the circle of search and significantly reduce the time to identify a specific person.
The system is based on a convolutional neural network (convolutional neural network, CNN). It is a subspecies of neural networks based on technologies of deep machine learning. CNN uses some features of the functioning of visual cortex. If you try to explain it in simple terms, there are segments that respond to simple signals (eg red) is a more complex conglomeration of simple functions (for example, all kinds of shirts). Many small segments may be part of several large (red can shirts, t-shirts, pants and so on). On the relationship between segments and the neural network can infer the presence of certain objects and their properties.
As for the algorithm, at the moment its accuracy is about 60% (on average, the neural network correctly guesses of 28 people from 41). This may seem enough, but this is only the first version of the algorithm that will improve. As stated by the developers
“Currently, the AI looks only at the growth, the color of his clothes and the floor. But nothing prevents to add more features. In addition, This experimental algorithm is ideal for cases when we need to collect data about a particular person. We already know its parameters. In this case, the AI will run very accurately. And when you consider that a couple of days in only one quarter of the camera and shoot about a million hours of video, the task becomes impossible for a human, but the algorithm can be several hundred times to reduce the number of hours of video to study.”
This and other news you can discuss in our chat in Telegram.