EN FR
• Legal notice
• Accessibility - non conforme

## Section: New Results

### Mobile Phone Sensing Middleware for Urban Pollution Monitoring

Participants : Valerie Issarny, Cong Kinh Nguyen, Pierre-Guillaume Raverdy, Fadwa Rebhi.

Mobile Phone Sensing (MPS) is a powerful solution for massive-scale sensing at low cost. The ubiquity of phones together with the rich set of sensors that they increasingly embed make mobile phones the devices of choice to sense our environment. Further, thanks to the – even sometimes unconscious – participation of people, MPS allows for leveraging both quantitative and qualitative sensing. And, still thanks to the participation of people who are moving across space, mobile phones may conveniently act as opportunistic proxies for the sensors in their communication range, which includes the fast developing wearables.

However, despite the numerous research work since the end 2000s, MPS keeps raising key challenges among which: How to make MPS resource-efficient$?$ How to mitigate mobile sensing heterogeneities$?$ How to involve and leverage the crowd$?$ How to leverage prior experiences$?$

Addressing the above MPS challenges primarily lies in taming the high heterogeneity not only of the computing system but also the crowd. The latter introduces a new dimension compared to traditional middleware research that has been concentrating on overcoming the heterogeneities of the computing infrastructure. In order to tackle these two dimensions together, we have been conducting a large scale empirical study in cooperation with the city of Paris (see http://tinyurl.com/soundcity-paris). Our experiment revolves around the public release of a MPS app for noise pollution monitoring that is built upon our dedicated mobile crowd-sensing middleware. Building on the Paris experiment, we systematically studied the influence of resource-efficiency and sensing accuracy on the effectiveness of the crowd participation [18]. In a complementary way, we analyzed user participation across time, so as to derive participation patterns that MPS middleware and application design may leverage.

Key take-away for MPS middleware and application design following our analysis includes:

• While contributors exhibit high heterogeneity regarding the accuracy of their sensors, they overall exhibit similar patterns. Location accuracy leads to discard about 60% of the observations and most observations are in the $\left[20-50\right]$ meters accuracy range. Noise sensing accuracy varies but calibration may be achieved per model rather than per device; calibration may then combine a number of techniques from comparison using a high-quality reference sensor to automated techniques leveraging assimilation and machine learning. Although our experiment is focused on noise sensing, we may expect similar results for other physical sensors. Overall, MPS allows collecting and assimilating relevant observations/measures. Still, the number of contributed measures by the MPS system needs to be high enough to overcome the low accuracy of the phone sensors.

• Although not specifically related to heterogeneity, energy efficiency is critical for the adoption of MPS. Our study confirms that energy-delay tradeoffs is a valuable approach; hence, the middleware must enable the buffering of the observations while the frequency of the transfers must be tuned by the application. Still, we notice that 30% of the observations reach the server after 2 hours even when observations are not buffered and are sent every 5mns, which indicates long periods of disconnection. Hence, if the timeliness of the observation is critical, then participatory sensing is most likely the approach to follow to ensure that the user is conscious about the sensing and activates appropriate network connection.

• The heterogeneity of the contributing crowd is obvious. However, it turns out to be an asset rather than a shortcoming of MPS. Indeed, the crowd overall exhibits similar contribution patterns across time. However, in the detail, each individual has different contribution patterns. This allows for the collection of complementary contributions over the whole day.

• The users appear to be still most of the time, while the user's activity cannot be qualified for 20% of the observations. This should be accounted for in the design of mobility-dependent MPS.

• One design issue that arises for MPS is whether to promote participatory or opportunistic sensing. It is our belief that a system (and thus supporting app) must support both. This enables to collect as many observations as possible from a large diversity of people, while participatory sensing guarantees contributions of higher quality.