Mobile Sensing: CarTel and Nericell

CarTel: A Distributed Mobile Sensor Computing System
Hull et al.

Nericell: Rich Monitoring of Road and Traffic Conditions using Mobile Smartphones
Mohan et al.

Today, we covered two papers on mobile wireless sensor networks. The
main idea behind mobile sensor networks is to be able to either
dynamically deploy sensors to a particular geographic region of
interest or to cover large geographic areas, where static sensor
installation costs would be prohibitive. The general thrust of this
avenue of research is supported by the increasing penetration of
commodity sensing devices on mobile platforms such as smartphones and
even x86 class hand-held computers (in contrast, most WSN papers that
talk about mote-class devices).

The first of the papers presented CarTel, a distributed, mobile sensor
network and telematics system with an integrated query engine. We
felt that the major contribution of this work was the implementation
of an actual working mobile sensor system, even though it was on a
small scale. However, the system did raise a few interesting issues.
The legality and morality of "borrowing" open wireless access points
by mobile CarTel nodes is questionable, even as the number of open
access points has declined in the last few years. ISPs servicing
last-mile Internet connections to consumers are now providing wireless
APs that are locked out-of-the-box, a strong indication that they will
continue to pursue enforcement of their terms-of-use prohibitions
against public connection sharing. This trend undermines the CarTel
premise (in 2006) that open access points would reach a density that
would allow CarTel to be deployed on a massive scale.

Since 2006, deployment of cellular 3G networks has become commonplace
and most current-generation smartphones and even some hand-held
computers now come with 3G capabilities. This wide-spread
availability of 3G obviates much of the motivation behind CarTel.
True, the data muling aspect of CarTel could be useful for places
where this kind of cellular infrastructure is unavailable. However,
it is getting more difficult to find such places. (One might --
incorrectly -- assume that such scenarios would arise in third-world
countries. However, it is in these countries that cell phone
penetration is even higher than computer penetration.)

We had a sense of slight unease about the way that CarTel incorporated
delay tolerance. Because data-muling and opportunistic open AP usage
are not predictable, data arriving at the base station ("portal" in
CarTel parlance) can be subject to large delays. So how "tolerant"
are we to delays? If delays become huge, one might as well just
purposefully drive a sensor-laden car to a particular region of
interest and drive back! The authors don't give an models of
mobility, node density, range, etc. and so can't provide and feel for
the bounds on delay.

We next turned our attention to several of CarTel's query language
constructs. First, the PRIORITY/DELIVERY ORDER and the
SUMMARIZE AS constructs caught our attention. Prioritization via the
end-point bisection method they propose implies a couple things (1)
that users are generally interested in a "big-picture" view of the
query results, and (2) that FIFO is bad because you can get a huge
delay in results. These are interesting points, but the
implementation of data priority implies in an idiosyncratic way that
real-time querying is the goal of CarTel, which is at odds with the
stated goal of delay tolerance. Second, we discussed the relative
merits of using the "trace" as the base unit of data. It was not
clear to us how GROUP BY a location would be implemented, for example.

We then switched focus to Nericell, a system for annotating road and
traffic conditions using sensor data from smartphones. The crux of
this paper is an interesting way to virtually re-orient sensors in 3D
to a canonical set of axes. Given this capability, simple heuristics
were used to detect potholes and traffic congestion (via accelerometer
and audio samples). We thought that the simple heuristics was a
weakness in this paper. More sophisticated statistical learning
methods exist for this kind of classification/discrimination.
Moreover, such methods may be orientation invariant, obviating the
need for the virtual orientation correction method on which Nericell
is based. On the other hand, the choice of smartphones is an astute
one, since cell phone penetration in India is quite high. Overall, we
felt that the major contribution of this paper is the real-world
implementation on unreliable/uncalibrated sensors, in an extreme
environment (Bangalore roads).