Counting People With WiFi
In the News: Gizmag, Huffington Post UK, PC World, ACM Tech News, Gizmodo, Science Daily, International Business Times UK, Tech Radar, Tech Times, Popular Mechanics, Phys.org, RT, UCSB press release, and other outlets
WiFi signals are everywhere these days. So how much information do they carry about us? For instance, can a WiFi link count the number of people walking in an area, without them carrying any device? Can this be done with only WiFi power measurements? In this project, we have shown how to do this. See the video and the paper for more details and results. |
Advisor: Yasamin Mostofi
Graduate Students: Saandeep Depatla and Arjun Muralidharan
Patent: SYSTEM AND METHOD OF OCCUPANCY ESTIMATION UTILIZING TRANSMITTED SIGNALS," patent # 20160294492, 2017
S. Depatla, A. Muralidharan, and Y. Mostofi, "Occupancy Estimation Using Only WiFi Power Measurements," IEEE Journal on Selected Areas in Communications (JSAC), special issue on Location-Awareness for Radios and Networks, 2015.[pdf]
We have proposed an approach for counting the number of people walking in an area, based on only the received power measurements of one WiFi link. The figures below show two sample outdoor and indoor scenarios, where a pair of WiFi cards are used for counting. The receiver card is recording its received power (RSSI) for a small period of time. We are then interested in estimating the total number of people based on only these received power measurements. Note that the robots are merely carrying the cards in this project.
|
People mainly impact the link in two ways. First, as a person crosses the Line of Sight (LOS), the received power drops considerably. Second, when she is not crossing the LOS, she acts as a scatterer, reflecting the signal and contributing to multipath fading. While these phenomena occur in any wave propagation scenario, the main challenge in this problem is finding a way to properly relate these to the number of people, which is what we have proposed here.
More specifically, through a Markov chain modeling, we have characterized the probability that any number of people cross the LOS. Furthermore, we have utilized the well-known K-distribution to characterize the scattering impact of one person. After some derivations, we finally derive an expression for the probability density function (pdf) of the received power as a function of the number of people present. During the experiment, the receiver WiFi card records its power measurements for a small period of time, from which we can get the experimental pdf. By using KL divergence as a metric, we can then estimate the number of people by comparing the experimental pdf to our derived theoretical one.
Here, we show extensive experimental results where a pair of WiFi cards (TX: D-link WBR1310 wireless router and RX: Atheros ar5006x card) are used for counting up to and including 9 people. We show the performance in both outdoor and indoor environments (shown above), as well as with both directional and omnidirectional antennas. In the directional case, an external directional antenna is connected to the WiFi card to limit multipath fading (see the paper for more on this). In the omnidirectional case, the WiFi card and its antenna are used as is (shown in Fig. 1).
Note that the video shows the results with the omnidirectional antennas that come with the WiFi cards.
Performance in Outdoors: The top tables show sample performances with (left) directional antennas and (right) omnidirectional antennas in the outdoor environment shown earlier in Figure 1 (left). It can be seen that the total number of people can be estimated well in both cases. The bottom figures show the CDF of the estimation error based on several runs for both cases. It can be seen that the estimation error is 2 or less 100% of the times for the directional case and 96% of the times for the omnidirectional case. |
Performance in Indoors: The top tables show sample performances with (left) directional antennas and (right) omnidirectional antennas in the indoor environment shown earlier in Figure 1 (right). It can be seen that the total number of people can be estimated well although indoor environments present more challenges. The bottom figures show the CDF of the estimation error based on several runs for both cases. It can be seen that the estimation error is 2 or less 100% of the times for the directional case and 63% of the times for the omnidirectional case. |
There are several potential applications that can benefit from an estimation of how crowded an area is. For instance, heating and cooling of a building can be better optimized based on learning the concentration of the people over the building. Emergency evacuation can also benefit from an estimation of the level of occupancy. Finally, stores can benefit from counting the number of shoppers for better business planning.
Given that WiFi networks are available in many buildings, we envision that they can provide a new way for occupancy estimation, in addition to cameras and other sensing mechanisms. In particular, its potential for counting behind walls can be a nice complement to existing vision-based methods.