AI uses WiFi data to estimate the number of people in a room

You may not know, but WiFi data can be applied to many interesting purposes. In addition to basic information such as signal strength, connection speed, or security level, WiFi data can also tell us quite a lot of information about connected objects, such as in the determination. See how many people are standing near a specific access point. In a new paper published recently with the title: 'DeepCount: Crowd Counting with WiFi via Deep Learning' - roughly translated: Estimate the number of people in a crowd via WiFi data, posted on the website Arxiv.org has succeeded in developing AI - DeepCount operational identity model - which helps estimate the number of people in a room from wireless data.

1. 91% of technology managers believe that AI will be the center of the next technology revolution

The work was carried out shortly after researchers at Ryerson University in Toronto announced a neural network that could help determine whether the owner of a smartphone is walking, cycling or driving. around some areas of the city using WiFi data, and at the same time another Purdue University study has also succeeded in developing a system that uses WiFi access logs for relationship analysis. Contacts between users, their locations and activities.

In this latest study, scientists took advantage of channel state information (CSI) - specifically phase and amplitude - to create a system of two artificial intelligence models, Includes an active identification model and deep learning model. The deep learning model is tasked with assessing the correlation between the number of people and channels by mapping their activities to CSI, while the operational identity model will be responsible for recording information when Someone walks in or out of the room through an 'electronic switch'. In the case of two models that obtain incompatible data - for example, if the operational identity model records a higher number of people than the deep learning model - DeepCount will use that difference to retrain the model. deep learning.

1. McAfee expert explained how deepfake and AI are drilling through the cyber security wall

In addition, the researchers have compiled a set of 800 CSI samples from 10 volunteers to participate in a variety of tasks, including activities such as waving, typing, sitting down. , walking, talking and eating (about 80% of the sample in each class is used for training purposes and the rest is used as a test kit). To train the operational identity model, scientists must first conduct amplitude data processing to eliminate noise and unwanted elements, then extract feature information. Deep-learning training is similar to the preprocessing stage, but is done with out-of-band phase data.

DeepCount running on a laptop has three receiving antennas, modified to report channel state data, and they are also connected to routers and two other transmitting antennas. Both operate on the 5GHz band to be able to produce wavelengths short enough to ensure better resolution, and also to minimize the ability to intervene from unwanted elements.

1. Google and DeepMind apply AI to predict the output of wind farms

In the experiments conducted, the research team reported that this deep learning model achieved an accuracy of up to 86.4% with up to 5 people. In addition, with the retraining of the models provided by the operational identity model, it has achieved accuracy in anticipated situations up to 90%.

'Our approach can show the' acceptable 'level of AI in determining the number of people via WiFi data in the context of complex changes in the environment within a throne. home. In theory, if it is possible to take into account the full range of indoor environments and use them as models to build a stronger model on a larger scale, we can fully apply this technology in The determination of the number of objects and objects in a much wider range, 'scientists said.