De draadloze én tagloze mensenmassameter

Dit onderzoeksproject ontwikkelt een tagloze techniek die in staat is om de grootte van mensenmassa’s in een bepaalde omgeving in te schatten.

transceivernode‘Tagloos’ betekent dat hierbij géén gebruik wordt gemaakt van toestellen die de mensen dragen of bij zich hebben (bijvoorbeeld smartphones, speciale festivalbandjes,…). Een klassieke aanpak voor dergelijke toepassingen is het gebruik van camerabeelden. Dit brengt echter privacy-issues met zich mee. Bovendien dient de omgeving op een vrij constante manier belicht te worden én vereist het meestal veel rekenkracht om deze zware algoritmen uit te voeren.

Onze techniek maakt gebruik van de invloed die een mensenmassa heeft op RF-signalen. Wij installeren een netwerk van transceivernodes aan de randen van de omgeving waarbinnen we een inschatting willen maken. De signaalsterktes van de communicatielinks binnen dit netwerk worden zowel gemeten wanneer de omgeving leeg is als wanneer er een evenement aan de gang is. Uit de verschillen tussen de twee metingen kan afgeleid worden hoeveel mensen er zich ongeveer in de omgeving bevinden.

Tijdens onze experimenten plaatsten we tot nu toe maximaal een zestigtal nodes aan de randen van een stage in een (semi-indoor) festivalomgeving. Indien mogelijk plaatsten we de nodes achter niet-metalen wanden of onder de houten toog van snackstands op een hoogte van ongeveer 1 meter. Met uitzondering van een speciale grote controllernode werden alle nodes via batterijen van hun energie voorzien. De controllernode stuurt de communicatie in het netwerk aan en slaat de gemeten signaalsterktes op. Idealiter is deze – afhankelijk van de aanwezige infrastructuur – verbonden met een lokaal netwerk waarlangs wij data kunnen ontvangen en verwerken om zo real-time resultaten weer te geven.

Een grote moeilijkheid die steeds terugkeert bij onze experimenten is het ontbreken van een ‘ground truth’. Om heel precies de efficiëntie van het systeem te kunnen evalueren heeft men uiteraard nood aan geverifieerde gegevens over het werkelijke aantal mensen dat zich in de omgeving bevindt. Bij een eerder experiment losten we dit probleem op met behulp van een reeks camerabeelden van zéér lage kwaliteit waar wij toegang tot hadden. Deze beelden werden manueel geanalyseerd door een groep vrijwilligers en geclassificeerd in 6 verschillende categorieën (0 – 5) afhankelijk van de grootte van de mensenmassa. Op die manier konden we nagaan of ons automatisch systeem even nauwkeurig was als een groot aantal paar mensenogen dat langdurig camerabeelden bestudeert.

De resultaten hiervan waren positief. In meer dan 90% van de gevallen bedroeg het verschil in inschatting tussen ons systeem en de groep vrijwilligers maximaal één categorie. Een meer recent experiment vond plaats in een speciale omgeving waar mensen enkel toegang tot hadden indien zij een armbandje lieten scannen door festivalpersoneel. Dit leverde een zeer waardevolle ground truth op voor ons experiment en uit de eerste resultaten blijkt dat er een vrijwel lineair verband bestaat tussen onze metingen en het ‘tagged’ scansysteem.

Radio Tomographic Imaging: het tracken van individuen

Ons onderzoek rond het inschatten van mensenmassa’s is oorspronkelijk ontstaan als een zijtak van een onderzoeksproject over ‘Radio Tomographic Imaging’ (RTI). Met deze tagloze techniek kan men op veel kleinere schaal automatisch individuen lokaliseren en tracken. Net zoals bij de mensenmassameter wordt er gebruik gemaakt van de invloed van mensen op RF-signalen binnen een netwerk van transceivernodes. Het RTI-algoritme zal op basis van deze metingen een afbeelding creëren, die voor elke locatie in de omgeving weergeeft in welke mate een verzwakkende invloed aanwezig is. Plaatsen waar veel verzwakking aanwezig is, worden verondersteld de individuen te bevatten die we willen lokaliseren.

afbeelding RTI

Schematische weergave van een RTI-systeem

Device Free Localization (DFL)

Over the past decade, there has been an increasing interest in research regarding device-free localization (DFL) systems. Unlike most types of automatic localization, DFL does not require the entity to wear an active or passive hardware device (tag). Instead, the influence that the physical presence of an entity has on its environment is used to determine its location. Classic examples such as applications involving assisted living or police and military raids in which applying a tag to each entity would be difficult or even outright impossible, clearly illustrate the usefulness of this concept.

One possible approach to DFL consists of relying on the impact the presence of an entity has on the propagation of RF-signals throughout the environments. In these systems, RF-transmitters and receivers are not (solely) used for communication, but act as sensors themselves. The techniques utilizing this approach can be subdivided into different categories. Through-the-wall radar-like ultra-wideband systems are commercially available and are primarily marketed towards police and military related applications. They are expensive, however, and potentially suffer from accuracy loss at larger distances. Other possible RF-based DFL methods consist of installing a network of RF-transceivers in an environment and measuring the impact of an entity on the signal strengths of the transmissions between these transceivers. The use of such a wireless sensor network (WSN) has become well-known in recent years, with techniques like Radio Tomographic Imaging (RTI) and passive fingerprinting.

DFL can be considered to consist of three different aspects: detection, tracking and identification.

Detection is defined as the ability of the system to detect changes in the environment and establish how many entities are responsible. Tracking refers to the tracking of the positions of these entities. A velocity and a sequence of positions of an entity are estimated over a certain duration of time based on a set of location estimations. Finally, identification is defined as determining the identity of the entities. ‘Identity’ can refer to the size or shape of the entity or to an actual identity of a human individual. This aspect is usually regarded as an extremely difficult task for tagless localization in general (’the identification problem’), only feasible to achieve through the use of optical-based techniques. For some applications, this can even be seen as an important advantage, due to the fact that a total lack of identification capabilities means that no privacy- related concerns can arise.

Our DFL-research within IDLab is primarily focused on detection and tracking, although we investigate these two aspects in vastly different contexts. The use of WSN-based tagless detection is studied within the domain of large-scale crowd estimation, while our tracking research is focused on small-scale multi-frequency RTI setups.

Estimating Crowd Sizes

Quite a lot of research has been performed in recent years regarding the use of wireless sensor networks to construct systems which are entirely focused on the detection aspect in environments which contain a large amount of human individuals: crowd density estimation. The density of a crowd can be important information for a multitude of applications. Examples include but are not limited to traffic control, crowd and safety management during events and gauging the interest of consumers for different products at a trade fair. The classic approach to obtaining this information is to make use of an optical camera-based system. The accuracy of these systems can be rather dependent on the lighting conditions and they tend to require the availability of a large amount of computing power. Furthermore, the use of optical cameras raises the specter of privacy- related issues.

A WSN-based system could potentially solve these issues. In the current state-of-the-art, however, the crowds on which the experiments are being performed tend to contain no more than a few tens of people. In order to look into the possibility of using these techniques for real-life applications, larger crowd sizes are needed. It is precisely this aspect that we investigate. We deploy our WSNs  (which operate primarily on the sub-GHz 433 MHz and 868 MHz bands) at large-scale events such as popular music festivals. At these types of locations, hundreds or even thousands of people can be present at a single stage. This provides us with large amounts of useful data to aid us in developing a full-fledged RF-based tagles crowd estimation system.

Radio Tomographic Imaging (RTI)

Radio Tomographic Imaging is an RF-based tagless localization technique also makes use of a WSN. It attempts to determine which locations within the environment were most likely responsible for measured RSS-changes of the communication links. The basic algorithm takes as input a list of RSS-differences between the current measurement and an earlier calibration measurement when the environment did not contain any targets and creates an attenuation image. This image is a representation of the environment in which the value of each pixel indicates the average attenuation a communication link will experience when its line-of-sight traverses the corresponding location. Pixels with a high amount of attenuation are assumed to be more likely to contain locatable entities.

Several variations and improvements of the basic RTI-technique do exist. In variance based RTI (VRTI), a windowed variance of the most recent RSS-link measurements is used as input to the algorithm. This eliminates the need for an earlier calibration step and paves the way for the use of RTI in emergency applications. Multi-channel approaches where the nodes communicate on multiple frequency channels have also been introduced in the state-of-the-art, as well as advanced multi-tracking systems which allow for accurate simultaneous tracking of up to 4 individuals.

In the majority of current Radio Tomographic Imaging systems, the nodes communicate using 2.4 GHz signals. Our research is primarily focused on investigating the potential use of multiple sub-GHz frequency bands in a single RTI-system and whether this approach could potentially solve several shortcomings of the technique. The larger wavelengths of sub-GHz signals makes it more difficult for them to be influenced by human presence, thus one would generally expect their use to consistently lead to worse localization accuracies. However, this same aspect also tends to make them less susceptible to multi-path effects caused by a multitude of static objects, which the RTI-algorithm interprets as noise. This can be an important advantage in complex environments. Furthermore, the increased communication range offered by sub-1 GHz frequencies could play an important role regarding the size of the environments in which a system can be installed. Finally, the use of sub-1 GHz frequencies could enable the creation of more energy-efficient RTI-systems using existing low-power communication solutions.

Projects

  • DOCPRO Stijn Denis [TBC]
  • ICON iFEST - improved festival experience through wireless
  • FWO SB Device-free crowd sensing at large music festivals using radio frequency signal features (submitted)

Key Publications

  • S. Denis, R. Berkvens, G. Ergeerts, B. Bellekens, and M. Weyn, “Combining multiple sub-1 GHz frequencies in Radio Tomographic Imaging,” in 2016 International Conference on Indoor Positioning and Indoor Navigation (IPIN), 2016, pp. 1–8.
  • S. Denis, R. Berkvens, G. Ergeerts, and M. Weyn, “Multi-frequency sub-1 GHz radio tomographic imaging in a complex indoor environment,” in 2017 International Conference on Indoor Positioning and Indoor Navigation (IPIN), 2017, pp. 1–8.
  • S. Denis, R. Berkvens, B. Bellekens, and M. Weyn, “Large Scale Crowd Density Estimation Using a sub-GHz Wireless Sensor Network,” in 2018 IEEE 29th Annual International Symposium on Personal, Indoor and Mobile Radio Communications (PIMRC), 2018, pp. 849–855.
Staff

Involved faculty

Maarten Weyn

Researchers

Contact

The Beacon
Sint-Pietersvliet 7
2000 Antwerpen
Belgium
DFL.IDLab@uantwerpen.be