Title: Towards Cooperative Localization of Wearable Sensors using Accelerometers and Cameras
1Towards Cooperative Localization of Wearable
Sensors using Accelerometers and Cameras
- Deokwoo Jung
- (with Thiago Teixeira and Andreas Savvides)
- deokwoo.jung_at_yale.edu
- Embedded Networks Applications Lab
- Yale University
2Indoor Localization
- Indoor localization is an essential technology
for many applications - Security Application, Assisted Living, Life
logging System etc - Indoor localization comparison
Cricket Priyantha.et.al ,00,04 RADAR Bahl.et.al,00 Surroundsense Azizyan.et.al,09 Our System
Precision Physical Location (cm) Physical Location (lt5 meter) Logical Location (gt5 meters) Physical Location (cm)
Mobile device Customized Sensor Node WLAN Card - Laptop Mobile Phone Mobile Phone/ Wearable Sensor
Infra- structure Ultrasound Beacon Nodes on Ceiling WLAN APs RF fingerprint database GSM Network Ambient signal database Networked cameras
Sensing modality Ultrasound RF signal Ambient signals- Light, sound, color Human walking motions
3Cooperative Localization Approach
- Our Approach
- Localizing mobile phones by combining their
built-in accelerometer (human motion) and
infrastructure camera (human centroids) - Why Human Centroids and Human Motion ?
- They are Complementary to each other
Wearable inertial sensors (Human Motion) Camera (Human Centroid)
ID tracking Accurate - node address Difficult feature extraction
Location Positioning Difficult walking orientation and distance estimation Accurate background subtraction
4Sensing Modeling and Approach
- Human Walking Model
- ? ?Accel. ? Walking distance
- The law of movement of human body by complex
kinetics - Inverted pendulum model of human gait.
- The body center of mass (BCOM) oscillates in the
z direction - as the person moves forward (y direction).
5 Statistical Analysis of Sensor Data
- Intuition BCOM follows a sinusoidal pattern
- Velocity of Body a Standard Deviation of Vertical
Acceleration - A correlation coefficient for the similarity
measure between accelerometer and camera data - Experiment
6Tracking Algorithm
- A camera extracts only centroid information
- Privacy Preserving and Cheap
- A simple tracking algorithm computes a speed of
anonymous centroids - associates human centroids in consecutive frames
based on their distances. - Problem Many possible ambiguous associations
7Path Disambiguation Problem
- Path Ambiguity Problem in Human Centroid Tracking
- A tracker associates one object with more than
two objects in two consecutive image frames when
two or more objects come close to each other.
8Disambiguation Algorithm
- Path Disambiguation as Non-linear Optimization
Problem - Find a set of association hypotheses to maximize
a matching rate, - The number of correct ID matchings between
accelerometers and centroids - Develop a search algorithm in a tree structure
- A leaf node a hypothesis of path segmentations
- Three stage pruning algorithm
- Sub-tree evaluation,
- Classification and Pruning,
- Reconstruction
9Clustering and Pruning in Hypothesis Tree
- Hypothesis Quality Metric
- how credible a given path hypothesis is compared
to others? - Correlation Coefficient Distance metric
- D(?H) E(?, e0H) - E(?, e1H)
Wrong Hypothesis
Correct Hypothesis
10Clustering and Pruning in Hypothesis Tree
- Leaf Clustering, Pruning, and Path Reconstruction
- Clusters the leaf nodes into groups and prunes
the subset of groups with lower metric values. - When only one leaf is left, reconstructs the
matching sequence
11Performance Evaluation via Experiment Simulation
- Experiment Setup
- A ceiling mounted camera (12ft) with a Intel
iMote2 node - Computes the centroid position of a person, 15
times per second. - A wearable sensor node with an Analog Devices
ADXL330 accelerometer on the persons waist - Collecting body acceleration data with 15Hz
sampling rate. - Transmitting its measurements to a computer
(fusion center) via a Zigbee wireless link. - People walk for 1 minute in a 5.4 m2 space.
12Experiment Dataset
- Walking trajectory of 12 people collected from
camera
13Similarity Metric Performance
- 100 matching rate without path ambiguity
14Disambiguation Algorithm Performance
- The performance depends on the level of crowd in
camera field of view. - Evaluate the performance using crowd density
metric, the number of pedestrians per area, m2,
Abishai.et.al, Pedestrian flow and level of
service - Crowd Density Scenario
Scenario A Normal flow B Restricted Flow C Dense Flow D Very Dense Flow
Crowd Density office building in business hour crowded shopping mall in weekend Crowded weekend party Subway station in Manhattan during the rush hour
People / m2 lt0.5 0.50.8 0.811.26 1.272
- If the crowd density gt 2, the pedestrian flow is
jammed, i.e. practically peoples movement
appears to be static - Our system is mainly targeting for the scenario A
(free flow), i.e. people can walk around without
much interaction.
15Performance over complexity of scenario
- The number of tracking errors grows with
polynomial order with crowd density (left) - The matching performance of disambiguation
algorithm for different crowd densities (right) - The performance gap is widening as crowd density
increases. - The performance becomes twice in the scenario D.
16Performance over disambiguation stages
- Matching rate improvement by disambiguation
algorithm
17Localization System Demonstration
- Controlled experiments with 10 people walking
scenario. - The performance of disambiguation algorithm
(right) is compared to the tracker-only
localization (left).
18Conclusion
- We presented a hybrid localization system using
accelerometers and cameras. - The proposed disambiguation algorithm operates
reliably, degrading gracefully even for crowded
scenarios - The constraint of accelerometer position (waist)
can be relaxed using additional inertial
measurement sensors. - Future work is to have a complete system
implementation running on a mobile phone More
sensors
19QUESTION ?
Thanks for your interest!
For more information, please visit
http//pantheon.yale.edu/dj92/