Skip Header NavigationIntranet 
CENTER FOR EMBEDDED NETWORKED SENSINGContactDirectionsEmploymentEventsNews
HomeAbout UsResearchEducationResourcesPeople

Research Project

FIND

Applications > Urban Sensing > FIND

OverviewResearch ProjectsPeopleResources

Software for Mobile Phone Based Sensing

In order for community members to participate in a sensing campaign, there must be a consistent and user-friendly manner in which these campaigns are run. Campaignr provides such a platform as a configurable, multi-modal, sensing application that can be deployed on Symbian OS 3rd Edition S60 smartphones. Our approach taken in the design of this system has been to be as general-purpose and flexible as possible, while still maintaining ease of use for the participants. A campaign is completely configurable via an XML configuration file for Campaignr. As a campaign creator, one must only be familiar with XML, as opposed to customized embedded-device programming languages and idioms. The Campaignr platform allows for the collection of location, image and audio data, among other modalities. These data can be collected at an automatic preset interval or at the command of the participant. Campaignr stores the collected data locally until it successfully uploads it to a pre-defined datastore location. Upload can occur concurrently with data collection, or at later time, depending on the needs of the specific campaign (e.g. reduced power consumption, poor network connectivity).

Applications

People

Modeling Location and Mobility

The mobility pattern and current location of mobile users in our systems are critical variables for various decisions relating to initial recruitment and subsequent tasking made by the network services in our application model. A proper understanding of these variables is therefore crucial, and is our focus in this activity. We analyzed a public database of location traces and found that a typical user can be summarized in terms of his main path and detours around the main path, which we modeled using a Bayesian sampling approach. Map-matching was an additional technique that was used in this approach.

People

overlay of location traces

Context and Activity Inferencing

A) Fine-grained
Fine-grained transportation mode information, such as whether a user is still, walking, running, biking, or in motorized transport, is important in many participatory urban sensing applications for the purpose of feedback. To this end, our research concentrated on creating an unobtrusive, orientation/position agnostic activity classification system that identifies the transportation mode of a user by using accelerometer and GPS sensors. Since our activities are kinetic based (motion dependent), we employ an accelerometer and a GPS as our sensors in a Nokia N95. We chose the following features for classification: Variance, Energy, FFT (1-5Hz) of Accelerometer and Speed of GPS. The classification system involved a Decision Tree in stage 1 and an HMM that incorporates history and transition probabilities for temporal smoothing in stage 2. The accuracy of this system, using a data set of six users with over 20 hours of data, was over 90 percent.

Activity Classification System
Activity Classification System

Speed and Accelerometer Data from Different Transportation Modes
Speed and Accelerometer Data from Different Transportation Modes

B) Coarse-grained
In another activity relating to context and activity classification, we focused on privacy and energy efficiency issues. Speed and map matching techniques are used to improve the interpretation of GPS location data and to improve the performance of activity classification. We developed an activity classifier and have achieved over 90% accuracy. However, GPS-based mobility characterization raises several issues related to spotty coverage, battery drain, and privacy threats.

For some applications and usage models we contend that it is desirable to adopt a more parsimoniousapproach to mobility characterization; one that avoids the collection and use of fine-grained location information by relying instead on GSM and WiFi connectivity data. Four features, number of unique cell IDs, residence time in a cell footprint, signal strength variance of WiFi data, and duration of dominant WiFi access point in view, are used to build a model and decision tree technique is used.

People

Security & Privacy Architecture

We build a Kerberos and Public Private key enabled security system for the Campaignr-based system. It is tailored for use with pervasive, widely available, and mobile sensor devices such as cell phones. Third-party applications can access the data using a simplistic access control schema. The system has build-in support for single-sign on and easy application frontend integration. The system is designed to use different encryption keys for handling of storage and transmission.

flow diagram of the security subsystem in campaignr

Location Privacy

We implemented two countermeasure methods: spatial rounding and noise addition, and evaluated the methods by measuring how much data corruption is necessary to hide frequent location information by counting correctly identified frequent locations. In spatial rounding, if the location data is too coarse, it will not correspond to the locations where users stayed. As expected, the number of correct addresses found decreases with increasing grid size. Spatial rounding with up to 1 meter x 1 meter grid identifies nearly same addresses. At least 1km x 1km grid is required to hide the all addresses, which is probably too coarse to be used many applications. In noise addition, we make use of the fact that if location data is noisy, it will not be useful for inferring the actual frequent location. The number of correctly identified addresses became nearly zero when the noise with 100 meter standard deviation were added. 100 meter noise may not change the output of application, however, it is doubtful that this can protect people’s privacy.

bar charts - preserving location privacy – effect of spatial rounding and noise on inference of frequent locations

People

Remote Fingerprinting of Participatory Sensing Devices

We investigated whether techniques proposed by Kohno et. al. (“Remote physical device fingerprinting”) can be used to identify urban sensing devices i.e. smart phones and Internet Tablets by their clock skew fingerprints instead of MAC address which are easier to spoof. We found that it is not possible to use the clock skew as fingerprints for smart phones because of too much noise. The Internet Tablets have better clock and unique fingerprints.

Further, we also evaluated the effect of temperature on clock skew based on Murdoch’s work (“Hot or not: revealing hidden services by their clock skew,” ACM CCS 2006). We correlated the changes in clock skew of participatory sensing Internet tables, to the ambient temperature profile over time of the location where these devices are used for data collection at that time, and use that to verify the devices location. For instance if we want to verify that data was collected at the LA harbor for on a particular day we compare the temperature profile for that day at the LA harbor with the devices clock skew to check if it coincides. This does not guarantee that the location is accurate but does increase its credibility of the data if they do coincide.

People

Network Infrastructure for Application Deployment

With matching funding from Cisco, and in collaboration with California State Parks we are building a metropolitan scale Wi-Fi network in downtown Los Angeles and two regions of the UCLA campus as a testbed for urban participatory sensing. Our research is focused on a network fabric architecture that can (a) embed network-attested location and time in sensor readings, and (b) provide context resolution control, selective data sharing, policy-based privacy, and related security mechanisms. This testbed is expanding UCLA’s urban-sensing network into the community around the new Los Angeles State Historic Park, a 32-acre site directly adjacent to LAs downtown. The infrastructure is currently in place for the two UCLA sites (APs covering 209,000 m2 in two outdoor courtyards) and we plan to finish deployment in mid-August. Using Cisco hardware, the system is designed as a VPN-connected wireless mesh network. Each of the sites is equipped with a Cisco wireless location appliance that determines and tracks client location based on Wi-Fi signal strength and triangulation techniques. The system will support the Urban Sensing Participatory Campaign framework in which community constituents will use off-the-shelf cell phones to sense, record and share urban and cultural phenomenon.

People

Event Detection

Inferring transitions in user's physical context plays a key role in summarizing the collected sensing data. For example, an assisted recall system providing continuous image capture generates a large number of images and requires reducing redundant images. To this end, our research exploited additional sensing data to assist image clustering in order to selectively display representative images to the users. We have found that observing the surrounding Bluetooth devices can sense activities, interactions, or physical situations.Through a theoretical and empirical analysis of Bluetooth characteristics, we found that using less than two inquiry windows limit the Bluetooth sensing coverage in half. However, the coverage only depends on the inquiry duration regardless of the window size when using four or more windows. Inquires using 4 and 8 windows were able to sense 80% and 100% of the neighboring devices, respectively. We then developed two noble Bluetooth episode-clustering algorithms, BT grouping based on k-means and Sliding window inspired by sudden changes in Bluetooth neighbors during transitions. We experimentally benchmark the two algorithms by comparing the number of episode detected, accuracy of episode boundaries, and effects of varying sampling rates. Our results demonstrate that BT grouping incurs less accuracy errors and more robust against reduced sampling rate. Similarity score is also proposed as a method to classify known episodes.

People