You can tell a lot about people from their Wi-Fi connections — including, as it turns out, how many of them are standing near an access point. In a newly published research paper (“DeepCount: Crowd Counting with WiFi via Deep Learning“) on the preprint server Arxiv.org, scientists describe an AI activity recognition model — DeepCount — that infers the population size of rooms from wireless data.
Their work comes not long after researchers at Ryerson University in Toronto demonstrated a neural network that can determine whether smartphone owners are walking, biking, or driving around a few city blocks by using Wi-Fi data, and after Purdue University researchers developed a system that uses Wi-Fi access logs to suss out relationships among users, locations, and activities.
In this latest study, the team leveraged channel state information (CSI) — specifically phase and amplitude — to create a two-model system consisting of an activity recognition model and deep learning model. The deep learning model was tasked with correlating the number of people and channels by mapping those people’s activities to CSI, while the former recognized when someone entered or left the room via an electronic switch. If the two models’ population counts disagreed — if, for instance, the activity recognition model recorded a higher number than the deep learning model — DeepCount used the discrepancy to retrain the deep learning model.
The researchers compiled a data set of 800 CSI samples from 10 volunteers engaged in a range of tasks, including waving, typing, sitting down, walking, talking, and eating. (About 80 percent of the samples in each class were used in training, and the rest were used as the test set.) To train the activity-recognizing model, they first preprocessed amplitude data to remove unwanted noise and interference and then extracted feature information. Training the deep learning model similarly involved preprocessing, but with phase data in addition to amplitude.
DeepCount ran on a laptop with three receiving antennas modified to report channel state data, which was connected to a router with two transmitting antennas. Both operated on a 5GHz frequency band in order to “make the wavelength short enough to ensure better resolution,” the researchers say, and to reduce the possibility of interference.
In experiments, the authors report that the deep learning model achieved 86.4 percent accuracy with up to five people, and that with retraining on samples supplied by the activity recognition model, it managed to achieve up to 90 percent prediction accuracy.
“Our approach can show acceptable accuracy in the context of complex changes in the indoor environment, which means [it] works fairly robust,” the researchers write. “In theory, if we can take into account enough circumstances in the indoor environment and [use] these as samples to build a robust model, we can reuse the model for the same environment.”